xref: /freebsd/sys/contrib/openzfs/module/zfs/zil.c (revision 7ef62cebc2f965b0f640263e179276928885e33d)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or https://opensource.org/licenses/CDDL-1.0.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23  * Copyright (c) 2011, 2018 by Delphix. All rights reserved.
24  * Copyright (c) 2014 Integros [integros.com]
25  * Copyright (c) 2018 Datto Inc.
26  */
27 
28 /* Portions Copyright 2010 Robert Milkowski */
29 
30 #include <sys/zfs_context.h>
31 #include <sys/spa.h>
32 #include <sys/spa_impl.h>
33 #include <sys/dmu.h>
34 #include <sys/zap.h>
35 #include <sys/arc.h>
36 #include <sys/stat.h>
37 #include <sys/zil.h>
38 #include <sys/zil_impl.h>
39 #include <sys/dsl_dataset.h>
40 #include <sys/vdev_impl.h>
41 #include <sys/dmu_tx.h>
42 #include <sys/dsl_pool.h>
43 #include <sys/metaslab.h>
44 #include <sys/trace_zfs.h>
45 #include <sys/abd.h>
46 #include <sys/brt.h>
47 #include <sys/wmsum.h>
48 
49 /*
50  * The ZFS Intent Log (ZIL) saves "transaction records" (itxs) of system
51  * calls that change the file system. Each itx has enough information to
52  * be able to replay them after a system crash, power loss, or
53  * equivalent failure mode. These are stored in memory until either:
54  *
55  *   1. they are committed to the pool by the DMU transaction group
56  *      (txg), at which point they can be discarded; or
57  *   2. they are committed to the on-disk ZIL for the dataset being
58  *      modified (e.g. due to an fsync, O_DSYNC, or other synchronous
59  *      requirement).
60  *
61  * In the event of a crash or power loss, the itxs contained by each
62  * dataset's on-disk ZIL will be replayed when that dataset is first
63  * instantiated (e.g. if the dataset is a normal filesystem, when it is
64  * first mounted).
65  *
66  * As hinted at above, there is one ZIL per dataset (both the in-memory
67  * representation, and the on-disk representation). The on-disk format
68  * consists of 3 parts:
69  *
70  * 	- a single, per-dataset, ZIL header; which points to a chain of
71  * 	- zero or more ZIL blocks; each of which contains
72  * 	- zero or more ZIL records
73  *
74  * A ZIL record holds the information necessary to replay a single
75  * system call transaction. A ZIL block can hold many ZIL records, and
76  * the blocks are chained together, similarly to a singly linked list.
77  *
78  * Each ZIL block contains a block pointer (blkptr_t) to the next ZIL
79  * block in the chain, and the ZIL header points to the first block in
80  * the chain.
81  *
82  * Note, there is not a fixed place in the pool to hold these ZIL
83  * blocks; they are dynamically allocated and freed as needed from the
84  * blocks available on the pool, though they can be preferentially
85  * allocated from a dedicated "log" vdev.
86  */
87 
88 /*
89  * This controls the amount of time that a ZIL block (lwb) will remain
90  * "open" when it isn't "full", and it has a thread waiting for it to be
91  * committed to stable storage. Please refer to the zil_commit_waiter()
92  * function (and the comments within it) for more details.
93  */
94 static uint_t zfs_commit_timeout_pct = 5;
95 
96 /*
97  * Minimal time we care to delay commit waiting for more ZIL records.
98  * At least FreeBSD kernel can't sleep for less than 2us at its best.
99  * So requests to sleep for less then 5us is a waste of CPU time with
100  * a risk of significant log latency increase due to oversleep.
101  */
102 static uint64_t zil_min_commit_timeout = 5000;
103 
104 /*
105  * See zil.h for more information about these fields.
106  */
107 static zil_kstat_values_t zil_stats = {
108 	{ "zil_commit_count",			KSTAT_DATA_UINT64 },
109 	{ "zil_commit_writer_count",		KSTAT_DATA_UINT64 },
110 	{ "zil_itx_count",			KSTAT_DATA_UINT64 },
111 	{ "zil_itx_indirect_count",		KSTAT_DATA_UINT64 },
112 	{ "zil_itx_indirect_bytes",		KSTAT_DATA_UINT64 },
113 	{ "zil_itx_copied_count",		KSTAT_DATA_UINT64 },
114 	{ "zil_itx_copied_bytes",		KSTAT_DATA_UINT64 },
115 	{ "zil_itx_needcopy_count",		KSTAT_DATA_UINT64 },
116 	{ "zil_itx_needcopy_bytes",		KSTAT_DATA_UINT64 },
117 	{ "zil_itx_metaslab_normal_count",	KSTAT_DATA_UINT64 },
118 	{ "zil_itx_metaslab_normal_bytes",	KSTAT_DATA_UINT64 },
119 	{ "zil_itx_metaslab_normal_write",	KSTAT_DATA_UINT64 },
120 	{ "zil_itx_metaslab_normal_alloc",	KSTAT_DATA_UINT64 },
121 	{ "zil_itx_metaslab_slog_count",	KSTAT_DATA_UINT64 },
122 	{ "zil_itx_metaslab_slog_bytes",	KSTAT_DATA_UINT64 },
123 	{ "zil_itx_metaslab_slog_write",	KSTAT_DATA_UINT64 },
124 	{ "zil_itx_metaslab_slog_alloc",	KSTAT_DATA_UINT64 },
125 };
126 
127 static zil_sums_t zil_sums_global;
128 static kstat_t *zil_kstats_global;
129 
130 /*
131  * Disable intent logging replay.  This global ZIL switch affects all pools.
132  */
133 int zil_replay_disable = 0;
134 
135 /*
136  * Disable the DKIOCFLUSHWRITECACHE commands that are normally sent to
137  * the disk(s) by the ZIL after an LWB write has completed. Setting this
138  * will cause ZIL corruption on power loss if a volatile out-of-order
139  * write cache is enabled.
140  */
141 static int zil_nocacheflush = 0;
142 
143 /*
144  * Limit SLOG write size per commit executed with synchronous priority.
145  * Any writes above that will be executed with lower (asynchronous) priority
146  * to limit potential SLOG device abuse by single active ZIL writer.
147  */
148 static uint64_t zil_slog_bulk = 768 * 1024;
149 
150 static kmem_cache_t *zil_lwb_cache;
151 static kmem_cache_t *zil_zcw_cache;
152 
153 static void zil_lwb_commit(zilog_t *zilog, lwb_t *lwb, itx_t *itx);
154 static itx_t *zil_itx_clone(itx_t *oitx);
155 
156 static int
157 zil_bp_compare(const void *x1, const void *x2)
158 {
159 	const dva_t *dva1 = &((zil_bp_node_t *)x1)->zn_dva;
160 	const dva_t *dva2 = &((zil_bp_node_t *)x2)->zn_dva;
161 
162 	int cmp = TREE_CMP(DVA_GET_VDEV(dva1), DVA_GET_VDEV(dva2));
163 	if (likely(cmp))
164 		return (cmp);
165 
166 	return (TREE_CMP(DVA_GET_OFFSET(dva1), DVA_GET_OFFSET(dva2)));
167 }
168 
169 static void
170 zil_bp_tree_init(zilog_t *zilog)
171 {
172 	avl_create(&zilog->zl_bp_tree, zil_bp_compare,
173 	    sizeof (zil_bp_node_t), offsetof(zil_bp_node_t, zn_node));
174 }
175 
176 static void
177 zil_bp_tree_fini(zilog_t *zilog)
178 {
179 	avl_tree_t *t = &zilog->zl_bp_tree;
180 	zil_bp_node_t *zn;
181 	void *cookie = NULL;
182 
183 	while ((zn = avl_destroy_nodes(t, &cookie)) != NULL)
184 		kmem_free(zn, sizeof (zil_bp_node_t));
185 
186 	avl_destroy(t);
187 }
188 
189 int
190 zil_bp_tree_add(zilog_t *zilog, const blkptr_t *bp)
191 {
192 	avl_tree_t *t = &zilog->zl_bp_tree;
193 	const dva_t *dva;
194 	zil_bp_node_t *zn;
195 	avl_index_t where;
196 
197 	if (BP_IS_EMBEDDED(bp))
198 		return (0);
199 
200 	dva = BP_IDENTITY(bp);
201 
202 	if (avl_find(t, dva, &where) != NULL)
203 		return (SET_ERROR(EEXIST));
204 
205 	zn = kmem_alloc(sizeof (zil_bp_node_t), KM_SLEEP);
206 	zn->zn_dva = *dva;
207 	avl_insert(t, zn, where);
208 
209 	return (0);
210 }
211 
212 static zil_header_t *
213 zil_header_in_syncing_context(zilog_t *zilog)
214 {
215 	return ((zil_header_t *)zilog->zl_header);
216 }
217 
218 static void
219 zil_init_log_chain(zilog_t *zilog, blkptr_t *bp)
220 {
221 	zio_cksum_t *zc = &bp->blk_cksum;
222 
223 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_0],
224 	    sizeof (zc->zc_word[ZIL_ZC_GUID_0]));
225 	(void) random_get_pseudo_bytes((void *)&zc->zc_word[ZIL_ZC_GUID_1],
226 	    sizeof (zc->zc_word[ZIL_ZC_GUID_1]));
227 	zc->zc_word[ZIL_ZC_OBJSET] = dmu_objset_id(zilog->zl_os);
228 	zc->zc_word[ZIL_ZC_SEQ] = 1ULL;
229 }
230 
231 static int
232 zil_kstats_global_update(kstat_t *ksp, int rw)
233 {
234 	zil_kstat_values_t *zs = ksp->ks_data;
235 	ASSERT3P(&zil_stats, ==, zs);
236 
237 	if (rw == KSTAT_WRITE) {
238 		return (SET_ERROR(EACCES));
239 	}
240 
241 	zil_kstat_values_update(zs, &zil_sums_global);
242 
243 	return (0);
244 }
245 
246 /*
247  * Read a log block and make sure it's valid.
248  */
249 static int
250 zil_read_log_block(zilog_t *zilog, boolean_t decrypt, const blkptr_t *bp,
251     blkptr_t *nbp, char **begin, char **end, arc_buf_t **abuf)
252 {
253 	zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
254 	arc_flags_t aflags = ARC_FLAG_WAIT;
255 	zbookmark_phys_t zb;
256 	int error;
257 
258 	if (zilog->zl_header->zh_claim_txg == 0)
259 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
260 
261 	if (!(zilog->zl_header->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
262 		zio_flags |= ZIO_FLAG_SPECULATIVE;
263 
264 	if (!decrypt)
265 		zio_flags |= ZIO_FLAG_RAW;
266 
267 	SET_BOOKMARK(&zb, bp->blk_cksum.zc_word[ZIL_ZC_OBJSET],
268 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL, bp->blk_cksum.zc_word[ZIL_ZC_SEQ]);
269 
270 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func,
271 	    abuf, ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
272 
273 	if (error == 0) {
274 		zio_cksum_t cksum = bp->blk_cksum;
275 
276 		/*
277 		 * Validate the checksummed log block.
278 		 *
279 		 * Sequence numbers should be... sequential.  The checksum
280 		 * verifier for the next block should be bp's checksum plus 1.
281 		 *
282 		 * Also check the log chain linkage and size used.
283 		 */
284 		cksum.zc_word[ZIL_ZC_SEQ]++;
285 
286 		uint64_t size = BP_GET_LSIZE(bp);
287 		if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
288 			zil_chain_t *zilc = (*abuf)->b_data;
289 			char *lr = (char *)(zilc + 1);
290 
291 			if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
292 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
293 			    zilc->zc_nused < sizeof (*zilc) ||
294 			    zilc->zc_nused > size) {
295 				error = SET_ERROR(ECKSUM);
296 			} else {
297 				*begin = lr;
298 				*end = lr + zilc->zc_nused - sizeof (*zilc);
299 				*nbp = zilc->zc_next_blk;
300 			}
301 		} else {
302 			char *lr = (*abuf)->b_data;
303 			zil_chain_t *zilc = (zil_chain_t *)(lr + size) - 1;
304 
305 			if (memcmp(&cksum, &zilc->zc_next_blk.blk_cksum,
306 			    sizeof (cksum)) || BP_IS_HOLE(&zilc->zc_next_blk) ||
307 			    (zilc->zc_nused > (size - sizeof (*zilc)))) {
308 				error = SET_ERROR(ECKSUM);
309 			} else {
310 				*begin = lr;
311 				*end = lr + zilc->zc_nused;
312 				*nbp = zilc->zc_next_blk;
313 			}
314 		}
315 	}
316 
317 	return (error);
318 }
319 
320 /*
321  * Read a TX_WRITE log data block.
322  */
323 static int
324 zil_read_log_data(zilog_t *zilog, const lr_write_t *lr, void *wbuf)
325 {
326 	zio_flag_t zio_flags = ZIO_FLAG_CANFAIL;
327 	const blkptr_t *bp = &lr->lr_blkptr;
328 	arc_flags_t aflags = ARC_FLAG_WAIT;
329 	arc_buf_t *abuf = NULL;
330 	zbookmark_phys_t zb;
331 	int error;
332 
333 	if (BP_IS_HOLE(bp)) {
334 		if (wbuf != NULL)
335 			memset(wbuf, 0, MAX(BP_GET_LSIZE(bp), lr->lr_length));
336 		return (0);
337 	}
338 
339 	if (zilog->zl_header->zh_claim_txg == 0)
340 		zio_flags |= ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB;
341 
342 	/*
343 	 * If we are not using the resulting data, we are just checking that
344 	 * it hasn't been corrupted so we don't need to waste CPU time
345 	 * decompressing and decrypting it.
346 	 */
347 	if (wbuf == NULL)
348 		zio_flags |= ZIO_FLAG_RAW;
349 
350 	ASSERT3U(BP_GET_LSIZE(bp), !=, 0);
351 	SET_BOOKMARK(&zb, dmu_objset_id(zilog->zl_os), lr->lr_foid,
352 	    ZB_ZIL_LEVEL, lr->lr_offset / BP_GET_LSIZE(bp));
353 
354 	error = arc_read(NULL, zilog->zl_spa, bp, arc_getbuf_func, &abuf,
355 	    ZIO_PRIORITY_SYNC_READ, zio_flags, &aflags, &zb);
356 
357 	if (error == 0) {
358 		if (wbuf != NULL)
359 			memcpy(wbuf, abuf->b_data, arc_buf_size(abuf));
360 		arc_buf_destroy(abuf, &abuf);
361 	}
362 
363 	return (error);
364 }
365 
366 void
367 zil_sums_init(zil_sums_t *zs)
368 {
369 	wmsum_init(&zs->zil_commit_count, 0);
370 	wmsum_init(&zs->zil_commit_writer_count, 0);
371 	wmsum_init(&zs->zil_itx_count, 0);
372 	wmsum_init(&zs->zil_itx_indirect_count, 0);
373 	wmsum_init(&zs->zil_itx_indirect_bytes, 0);
374 	wmsum_init(&zs->zil_itx_copied_count, 0);
375 	wmsum_init(&zs->zil_itx_copied_bytes, 0);
376 	wmsum_init(&zs->zil_itx_needcopy_count, 0);
377 	wmsum_init(&zs->zil_itx_needcopy_bytes, 0);
378 	wmsum_init(&zs->zil_itx_metaslab_normal_count, 0);
379 	wmsum_init(&zs->zil_itx_metaslab_normal_bytes, 0);
380 	wmsum_init(&zs->zil_itx_metaslab_normal_write, 0);
381 	wmsum_init(&zs->zil_itx_metaslab_normal_alloc, 0);
382 	wmsum_init(&zs->zil_itx_metaslab_slog_count, 0);
383 	wmsum_init(&zs->zil_itx_metaslab_slog_bytes, 0);
384 	wmsum_init(&zs->zil_itx_metaslab_slog_write, 0);
385 	wmsum_init(&zs->zil_itx_metaslab_slog_alloc, 0);
386 }
387 
388 void
389 zil_sums_fini(zil_sums_t *zs)
390 {
391 	wmsum_fini(&zs->zil_commit_count);
392 	wmsum_fini(&zs->zil_commit_writer_count);
393 	wmsum_fini(&zs->zil_itx_count);
394 	wmsum_fini(&zs->zil_itx_indirect_count);
395 	wmsum_fini(&zs->zil_itx_indirect_bytes);
396 	wmsum_fini(&zs->zil_itx_copied_count);
397 	wmsum_fini(&zs->zil_itx_copied_bytes);
398 	wmsum_fini(&zs->zil_itx_needcopy_count);
399 	wmsum_fini(&zs->zil_itx_needcopy_bytes);
400 	wmsum_fini(&zs->zil_itx_metaslab_normal_count);
401 	wmsum_fini(&zs->zil_itx_metaslab_normal_bytes);
402 	wmsum_fini(&zs->zil_itx_metaslab_normal_write);
403 	wmsum_fini(&zs->zil_itx_metaslab_normal_alloc);
404 	wmsum_fini(&zs->zil_itx_metaslab_slog_count);
405 	wmsum_fini(&zs->zil_itx_metaslab_slog_bytes);
406 	wmsum_fini(&zs->zil_itx_metaslab_slog_write);
407 	wmsum_fini(&zs->zil_itx_metaslab_slog_alloc);
408 }
409 
410 void
411 zil_kstat_values_update(zil_kstat_values_t *zs, zil_sums_t *zil_sums)
412 {
413 	zs->zil_commit_count.value.ui64 =
414 	    wmsum_value(&zil_sums->zil_commit_count);
415 	zs->zil_commit_writer_count.value.ui64 =
416 	    wmsum_value(&zil_sums->zil_commit_writer_count);
417 	zs->zil_itx_count.value.ui64 =
418 	    wmsum_value(&zil_sums->zil_itx_count);
419 	zs->zil_itx_indirect_count.value.ui64 =
420 	    wmsum_value(&zil_sums->zil_itx_indirect_count);
421 	zs->zil_itx_indirect_bytes.value.ui64 =
422 	    wmsum_value(&zil_sums->zil_itx_indirect_bytes);
423 	zs->zil_itx_copied_count.value.ui64 =
424 	    wmsum_value(&zil_sums->zil_itx_copied_count);
425 	zs->zil_itx_copied_bytes.value.ui64 =
426 	    wmsum_value(&zil_sums->zil_itx_copied_bytes);
427 	zs->zil_itx_needcopy_count.value.ui64 =
428 	    wmsum_value(&zil_sums->zil_itx_needcopy_count);
429 	zs->zil_itx_needcopy_bytes.value.ui64 =
430 	    wmsum_value(&zil_sums->zil_itx_needcopy_bytes);
431 	zs->zil_itx_metaslab_normal_count.value.ui64 =
432 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_count);
433 	zs->zil_itx_metaslab_normal_bytes.value.ui64 =
434 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_bytes);
435 	zs->zil_itx_metaslab_normal_write.value.ui64 =
436 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_write);
437 	zs->zil_itx_metaslab_normal_alloc.value.ui64 =
438 	    wmsum_value(&zil_sums->zil_itx_metaslab_normal_alloc);
439 	zs->zil_itx_metaslab_slog_count.value.ui64 =
440 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_count);
441 	zs->zil_itx_metaslab_slog_bytes.value.ui64 =
442 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_bytes);
443 	zs->zil_itx_metaslab_slog_write.value.ui64 =
444 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_write);
445 	zs->zil_itx_metaslab_slog_alloc.value.ui64 =
446 	    wmsum_value(&zil_sums->zil_itx_metaslab_slog_alloc);
447 }
448 
449 /*
450  * Parse the intent log, and call parse_func for each valid record within.
451  */
452 int
453 zil_parse(zilog_t *zilog, zil_parse_blk_func_t *parse_blk_func,
454     zil_parse_lr_func_t *parse_lr_func, void *arg, uint64_t txg,
455     boolean_t decrypt)
456 {
457 	const zil_header_t *zh = zilog->zl_header;
458 	boolean_t claimed = !!zh->zh_claim_txg;
459 	uint64_t claim_blk_seq = claimed ? zh->zh_claim_blk_seq : UINT64_MAX;
460 	uint64_t claim_lr_seq = claimed ? zh->zh_claim_lr_seq : UINT64_MAX;
461 	uint64_t max_blk_seq = 0;
462 	uint64_t max_lr_seq = 0;
463 	uint64_t blk_count = 0;
464 	uint64_t lr_count = 0;
465 	blkptr_t blk, next_blk = {{{{0}}}};
466 	int error = 0;
467 
468 	/*
469 	 * Old logs didn't record the maximum zh_claim_lr_seq.
470 	 */
471 	if (!(zh->zh_flags & ZIL_CLAIM_LR_SEQ_VALID))
472 		claim_lr_seq = UINT64_MAX;
473 
474 	/*
475 	 * Starting at the block pointed to by zh_log we read the log chain.
476 	 * For each block in the chain we strongly check that block to
477 	 * ensure its validity.  We stop when an invalid block is found.
478 	 * For each block pointer in the chain we call parse_blk_func().
479 	 * For each record in each valid block we call parse_lr_func().
480 	 * If the log has been claimed, stop if we encounter a sequence
481 	 * number greater than the highest claimed sequence number.
482 	 */
483 	zil_bp_tree_init(zilog);
484 
485 	for (blk = zh->zh_log; !BP_IS_HOLE(&blk); blk = next_blk) {
486 		uint64_t blk_seq = blk.blk_cksum.zc_word[ZIL_ZC_SEQ];
487 		int reclen;
488 		char *lrp, *end;
489 		arc_buf_t *abuf = NULL;
490 
491 		if (blk_seq > claim_blk_seq)
492 			break;
493 
494 		error = parse_blk_func(zilog, &blk, arg, txg);
495 		if (error != 0)
496 			break;
497 		ASSERT3U(max_blk_seq, <, blk_seq);
498 		max_blk_seq = blk_seq;
499 		blk_count++;
500 
501 		if (max_lr_seq == claim_lr_seq && max_blk_seq == claim_blk_seq)
502 			break;
503 
504 		error = zil_read_log_block(zilog, decrypt, &blk, &next_blk,
505 		    &lrp, &end, &abuf);
506 		if (error != 0) {
507 			if (abuf)
508 				arc_buf_destroy(abuf, &abuf);
509 			if (claimed) {
510 				char name[ZFS_MAX_DATASET_NAME_LEN];
511 
512 				dmu_objset_name(zilog->zl_os, name);
513 
514 				cmn_err(CE_WARN, "ZFS read log block error %d, "
515 				    "dataset %s, seq 0x%llx\n", error, name,
516 				    (u_longlong_t)blk_seq);
517 			}
518 			break;
519 		}
520 
521 		for (; lrp < end; lrp += reclen) {
522 			lr_t *lr = (lr_t *)lrp;
523 			reclen = lr->lrc_reclen;
524 			ASSERT3U(reclen, >=, sizeof (lr_t));
525 			if (lr->lrc_seq > claim_lr_seq) {
526 				arc_buf_destroy(abuf, &abuf);
527 				goto done;
528 			}
529 
530 			error = parse_lr_func(zilog, lr, arg, txg);
531 			if (error != 0) {
532 				arc_buf_destroy(abuf, &abuf);
533 				goto done;
534 			}
535 			ASSERT3U(max_lr_seq, <, lr->lrc_seq);
536 			max_lr_seq = lr->lrc_seq;
537 			lr_count++;
538 		}
539 		arc_buf_destroy(abuf, &abuf);
540 	}
541 done:
542 	zilog->zl_parse_error = error;
543 	zilog->zl_parse_blk_seq = max_blk_seq;
544 	zilog->zl_parse_lr_seq = max_lr_seq;
545 	zilog->zl_parse_blk_count = blk_count;
546 	zilog->zl_parse_lr_count = lr_count;
547 
548 	zil_bp_tree_fini(zilog);
549 
550 	return (error);
551 }
552 
553 static int
554 zil_clear_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
555     uint64_t first_txg)
556 {
557 	(void) tx;
558 	ASSERT(!BP_IS_HOLE(bp));
559 
560 	/*
561 	 * As we call this function from the context of a rewind to a
562 	 * checkpoint, each ZIL block whose txg is later than the txg
563 	 * that we rewind to is invalid. Thus, we return -1 so
564 	 * zil_parse() doesn't attempt to read it.
565 	 */
566 	if (bp->blk_birth >= first_txg)
567 		return (-1);
568 
569 	if (zil_bp_tree_add(zilog, bp) != 0)
570 		return (0);
571 
572 	zio_free(zilog->zl_spa, first_txg, bp);
573 	return (0);
574 }
575 
576 static int
577 zil_noop_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
578     uint64_t first_txg)
579 {
580 	(void) zilog, (void) lrc, (void) tx, (void) first_txg;
581 	return (0);
582 }
583 
584 static int
585 zil_claim_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
586     uint64_t first_txg)
587 {
588 	/*
589 	 * Claim log block if not already committed and not already claimed.
590 	 * If tx == NULL, just verify that the block is claimable.
591 	 */
592 	if (BP_IS_HOLE(bp) || bp->blk_birth < first_txg ||
593 	    zil_bp_tree_add(zilog, bp) != 0)
594 		return (0);
595 
596 	return (zio_wait(zio_claim(NULL, zilog->zl_spa,
597 	    tx == NULL ? 0 : first_txg, bp, spa_claim_notify, NULL,
598 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_SPECULATIVE | ZIO_FLAG_SCRUB)));
599 }
600 
601 static int
602 zil_claim_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t first_txg)
603 {
604 	lr_write_t *lr = (lr_write_t *)lrc;
605 	int error;
606 
607 	ASSERT(lrc->lrc_txtype == TX_WRITE);
608 
609 	/*
610 	 * If the block is not readable, don't claim it.  This can happen
611 	 * in normal operation when a log block is written to disk before
612 	 * some of the dmu_sync() blocks it points to.  In this case, the
613 	 * transaction cannot have been committed to anyone (we would have
614 	 * waited for all writes to be stable first), so it is semantically
615 	 * correct to declare this the end of the log.
616 	 */
617 	if (lr->lr_blkptr.blk_birth >= first_txg) {
618 		error = zil_read_log_data(zilog, lr, NULL);
619 		if (error != 0)
620 			return (error);
621 	}
622 
623 	return (zil_claim_log_block(zilog, &lr->lr_blkptr, tx, first_txg));
624 }
625 
626 static int
627 zil_claim_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx)
628 {
629 	const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc;
630 	const blkptr_t *bp;
631 	spa_t *spa;
632 	uint_t ii;
633 
634 	ASSERT(lrc->lrc_txtype == TX_CLONE_RANGE);
635 
636 	if (tx == NULL) {
637 		return (0);
638 	}
639 
640 	/*
641 	 * XXX: Do we need to byteswap lr?
642 	 */
643 
644 	spa = zilog->zl_spa;
645 
646 	for (ii = 0; ii < lr->lr_nbps; ii++) {
647 		bp = &lr->lr_bps[ii];
648 
649 		/*
650 		 * When data in embedded into BP there is no need to create
651 		 * BRT entry as there is no data block. Just copy the BP as
652 		 * it contains the data.
653 		 */
654 		if (!BP_IS_HOLE(bp) && !BP_IS_EMBEDDED(bp)) {
655 			brt_pending_add(spa, bp, tx);
656 		}
657 	}
658 
659 	return (0);
660 }
661 
662 static int
663 zil_claim_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
664     uint64_t first_txg)
665 {
666 
667 	switch (lrc->lrc_txtype) {
668 	case TX_WRITE:
669 		return (zil_claim_write(zilog, lrc, tx, first_txg));
670 	case TX_CLONE_RANGE:
671 		return (zil_claim_clone_range(zilog, lrc, tx));
672 	default:
673 		return (0);
674 	}
675 }
676 
677 static int
678 zil_free_log_block(zilog_t *zilog, const blkptr_t *bp, void *tx,
679     uint64_t claim_txg)
680 {
681 	(void) claim_txg;
682 
683 	zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
684 
685 	return (0);
686 }
687 
688 static int
689 zil_free_write(zilog_t *zilog, const lr_t *lrc, void *tx, uint64_t claim_txg)
690 {
691 	lr_write_t *lr = (lr_write_t *)lrc;
692 	blkptr_t *bp = &lr->lr_blkptr;
693 
694 	ASSERT(lrc->lrc_txtype == TX_WRITE);
695 
696 	/*
697 	 * If we previously claimed it, we need to free it.
698 	 */
699 	if (bp->blk_birth >= claim_txg && zil_bp_tree_add(zilog, bp) == 0 &&
700 	    !BP_IS_HOLE(bp)) {
701 		zio_free(zilog->zl_spa, dmu_tx_get_txg(tx), bp);
702 	}
703 
704 	return (0);
705 }
706 
707 static int
708 zil_free_clone_range(zilog_t *zilog, const lr_t *lrc, void *tx)
709 {
710 	const lr_clone_range_t *lr = (const lr_clone_range_t *)lrc;
711 	const blkptr_t *bp;
712 	spa_t *spa;
713 	uint_t ii;
714 
715 	ASSERT(lrc->lrc_txtype == TX_CLONE_RANGE);
716 
717 	if (tx == NULL) {
718 		return (0);
719 	}
720 
721 	spa = zilog->zl_spa;
722 
723 	for (ii = 0; ii < lr->lr_nbps; ii++) {
724 		bp = &lr->lr_bps[ii];
725 
726 		if (!BP_IS_HOLE(bp)) {
727 			zio_free(spa, dmu_tx_get_txg(tx), bp);
728 		}
729 	}
730 
731 	return (0);
732 }
733 
734 static int
735 zil_free_log_record(zilog_t *zilog, const lr_t *lrc, void *tx,
736     uint64_t claim_txg)
737 {
738 
739 	if (claim_txg == 0) {
740 		return (0);
741 	}
742 
743 	switch (lrc->lrc_txtype) {
744 	case TX_WRITE:
745 		return (zil_free_write(zilog, lrc, tx, claim_txg));
746 	case TX_CLONE_RANGE:
747 		return (zil_free_clone_range(zilog, lrc, tx));
748 	default:
749 		return (0);
750 	}
751 }
752 
753 static int
754 zil_lwb_vdev_compare(const void *x1, const void *x2)
755 {
756 	const uint64_t v1 = ((zil_vdev_node_t *)x1)->zv_vdev;
757 	const uint64_t v2 = ((zil_vdev_node_t *)x2)->zv_vdev;
758 
759 	return (TREE_CMP(v1, v2));
760 }
761 
762 static lwb_t *
763 zil_alloc_lwb(zilog_t *zilog, blkptr_t *bp, boolean_t slog, uint64_t txg,
764     boolean_t fastwrite)
765 {
766 	lwb_t *lwb;
767 
768 	lwb = kmem_cache_alloc(zil_lwb_cache, KM_SLEEP);
769 	lwb->lwb_zilog = zilog;
770 	lwb->lwb_blk = *bp;
771 	lwb->lwb_fastwrite = fastwrite;
772 	lwb->lwb_slog = slog;
773 	lwb->lwb_indirect = B_FALSE;
774 	if (BP_GET_CHECKSUM(bp) == ZIO_CHECKSUM_ZILOG2) {
775 		lwb->lwb_nused = lwb->lwb_nfilled = sizeof (zil_chain_t);
776 		lwb->lwb_sz = BP_GET_LSIZE(bp);
777 	} else {
778 		lwb->lwb_nused = lwb->lwb_nfilled = 0;
779 		lwb->lwb_sz = BP_GET_LSIZE(bp) - sizeof (zil_chain_t);
780 	}
781 	lwb->lwb_state = LWB_STATE_CLOSED;
782 	lwb->lwb_buf = zio_buf_alloc(BP_GET_LSIZE(bp));
783 	lwb->lwb_write_zio = NULL;
784 	lwb->lwb_root_zio = NULL;
785 	lwb->lwb_issued_timestamp = 0;
786 	lwb->lwb_issued_txg = 0;
787 	lwb->lwb_max_txg = txg;
788 
789 	mutex_enter(&zilog->zl_lock);
790 	list_insert_tail(&zilog->zl_lwb_list, lwb);
791 	mutex_exit(&zilog->zl_lock);
792 
793 	return (lwb);
794 }
795 
796 static void
797 zil_free_lwb(zilog_t *zilog, lwb_t *lwb)
798 {
799 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
800 	ASSERT(!MUTEX_HELD(&lwb->lwb_vdev_lock));
801 	VERIFY(list_is_empty(&lwb->lwb_waiters));
802 	VERIFY(list_is_empty(&lwb->lwb_itxs));
803 	ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
804 	ASSERT3P(lwb->lwb_write_zio, ==, NULL);
805 	ASSERT3P(lwb->lwb_root_zio, ==, NULL);
806 	ASSERT3U(lwb->lwb_max_txg, <=, spa_syncing_txg(zilog->zl_spa));
807 	ASSERT(lwb->lwb_state == LWB_STATE_CLOSED ||
808 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
809 
810 	/*
811 	 * Clear the zilog's field to indicate this lwb is no longer
812 	 * valid, and prevent use-after-free errors.
813 	 */
814 	if (zilog->zl_last_lwb_opened == lwb)
815 		zilog->zl_last_lwb_opened = NULL;
816 
817 	kmem_cache_free(zil_lwb_cache, lwb);
818 }
819 
820 /*
821  * Called when we create in-memory log transactions so that we know
822  * to cleanup the itxs at the end of spa_sync().
823  */
824 static void
825 zilog_dirty(zilog_t *zilog, uint64_t txg)
826 {
827 	dsl_pool_t *dp = zilog->zl_dmu_pool;
828 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
829 
830 	ASSERT(spa_writeable(zilog->zl_spa));
831 
832 	if (ds->ds_is_snapshot)
833 		panic("dirtying snapshot!");
834 
835 	if (txg_list_add(&dp->dp_dirty_zilogs, zilog, txg)) {
836 		/* up the hold count until we can be written out */
837 		dmu_buf_add_ref(ds->ds_dbuf, zilog);
838 
839 		zilog->zl_dirty_max_txg = MAX(txg, zilog->zl_dirty_max_txg);
840 	}
841 }
842 
843 /*
844  * Determine if the zil is dirty in the specified txg. Callers wanting to
845  * ensure that the dirty state does not change must hold the itxg_lock for
846  * the specified txg. Holding the lock will ensure that the zil cannot be
847  * dirtied (zil_itx_assign) or cleaned (zil_clean) while we check its current
848  * state.
849  */
850 static boolean_t __maybe_unused
851 zilog_is_dirty_in_txg(zilog_t *zilog, uint64_t txg)
852 {
853 	dsl_pool_t *dp = zilog->zl_dmu_pool;
854 
855 	if (txg_list_member(&dp->dp_dirty_zilogs, zilog, txg & TXG_MASK))
856 		return (B_TRUE);
857 	return (B_FALSE);
858 }
859 
860 /*
861  * Determine if the zil is dirty. The zil is considered dirty if it has
862  * any pending itx records that have not been cleaned by zil_clean().
863  */
864 static boolean_t
865 zilog_is_dirty(zilog_t *zilog)
866 {
867 	dsl_pool_t *dp = zilog->zl_dmu_pool;
868 
869 	for (int t = 0; t < TXG_SIZE; t++) {
870 		if (txg_list_member(&dp->dp_dirty_zilogs, zilog, t))
871 			return (B_TRUE);
872 	}
873 	return (B_FALSE);
874 }
875 
876 /*
877  * Its called in zil_commit context (zil_process_commit_list()/zil_create()).
878  * It activates SPA_FEATURE_ZILSAXATTR feature, if its enabled.
879  * Check dsl_dataset_feature_is_active to avoid txg_wait_synced() on every
880  * zil_commit.
881  */
882 static void
883 zil_commit_activate_saxattr_feature(zilog_t *zilog)
884 {
885 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
886 	uint64_t txg = 0;
887 	dmu_tx_t *tx = NULL;
888 
889 	if (spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) &&
890 	    dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL &&
891 	    !dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR)) {
892 		tx = dmu_tx_create(zilog->zl_os);
893 		VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
894 		dsl_dataset_dirty(ds, tx);
895 		txg = dmu_tx_get_txg(tx);
896 
897 		mutex_enter(&ds->ds_lock);
898 		ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
899 		    (void *)B_TRUE;
900 		mutex_exit(&ds->ds_lock);
901 		dmu_tx_commit(tx);
902 		txg_wait_synced(zilog->zl_dmu_pool, txg);
903 	}
904 }
905 
906 /*
907  * Create an on-disk intent log.
908  */
909 static lwb_t *
910 zil_create(zilog_t *zilog)
911 {
912 	const zil_header_t *zh = zilog->zl_header;
913 	lwb_t *lwb = NULL;
914 	uint64_t txg = 0;
915 	dmu_tx_t *tx = NULL;
916 	blkptr_t blk;
917 	int error = 0;
918 	boolean_t fastwrite = FALSE;
919 	boolean_t slog = FALSE;
920 	dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
921 
922 
923 	/*
924 	 * Wait for any previous destroy to complete.
925 	 */
926 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
927 
928 	ASSERT(zh->zh_claim_txg == 0);
929 	ASSERT(zh->zh_replay_seq == 0);
930 
931 	blk = zh->zh_log;
932 
933 	/*
934 	 * Allocate an initial log block if:
935 	 *    - there isn't one already
936 	 *    - the existing block is the wrong endianness
937 	 */
938 	if (BP_IS_HOLE(&blk) || BP_SHOULD_BYTESWAP(&blk)) {
939 		tx = dmu_tx_create(zilog->zl_os);
940 		VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
941 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
942 		txg = dmu_tx_get_txg(tx);
943 
944 		if (!BP_IS_HOLE(&blk)) {
945 			zio_free(zilog->zl_spa, txg, &blk);
946 			BP_ZERO(&blk);
947 		}
948 
949 		error = zio_alloc_zil(zilog->zl_spa, zilog->zl_os, txg, &blk,
950 		    ZIL_MIN_BLKSZ, &slog);
951 		fastwrite = TRUE;
952 
953 		if (error == 0)
954 			zil_init_log_chain(zilog, &blk);
955 	}
956 
957 	/*
958 	 * Allocate a log write block (lwb) for the first log block.
959 	 */
960 	if (error == 0)
961 		lwb = zil_alloc_lwb(zilog, &blk, slog, txg, fastwrite);
962 
963 	/*
964 	 * If we just allocated the first log block, commit our transaction
965 	 * and wait for zil_sync() to stuff the block pointer into zh_log.
966 	 * (zh is part of the MOS, so we cannot modify it in open context.)
967 	 */
968 	if (tx != NULL) {
969 		/*
970 		 * If "zilsaxattr" feature is enabled on zpool, then activate
971 		 * it now when we're creating the ZIL chain. We can't wait with
972 		 * this until we write the first xattr log record because we
973 		 * need to wait for the feature activation to sync out.
974 		 */
975 		if (spa_feature_is_enabled(zilog->zl_spa,
976 		    SPA_FEATURE_ZILSAXATTR) && dmu_objset_type(zilog->zl_os) !=
977 		    DMU_OST_ZVOL) {
978 			mutex_enter(&ds->ds_lock);
979 			ds->ds_feature_activation[SPA_FEATURE_ZILSAXATTR] =
980 			    (void *)B_TRUE;
981 			mutex_exit(&ds->ds_lock);
982 		}
983 
984 		dmu_tx_commit(tx);
985 		txg_wait_synced(zilog->zl_dmu_pool, txg);
986 	} else {
987 		/*
988 		 * This branch covers the case where we enable the feature on a
989 		 * zpool that has existing ZIL headers.
990 		 */
991 		zil_commit_activate_saxattr_feature(zilog);
992 	}
993 	IMPLY(spa_feature_is_enabled(zilog->zl_spa, SPA_FEATURE_ZILSAXATTR) &&
994 	    dmu_objset_type(zilog->zl_os) != DMU_OST_ZVOL,
995 	    dsl_dataset_feature_is_active(ds, SPA_FEATURE_ZILSAXATTR));
996 
997 	ASSERT(error != 0 || memcmp(&blk, &zh->zh_log, sizeof (blk)) == 0);
998 	IMPLY(error == 0, lwb != NULL);
999 
1000 	return (lwb);
1001 }
1002 
1003 /*
1004  * In one tx, free all log blocks and clear the log header. If keep_first
1005  * is set, then we're replaying a log with no content. We want to keep the
1006  * first block, however, so that the first synchronous transaction doesn't
1007  * require a txg_wait_synced() in zil_create(). We don't need to
1008  * txg_wait_synced() here either when keep_first is set, because both
1009  * zil_create() and zil_destroy() will wait for any in-progress destroys
1010  * to complete.
1011  * Return B_TRUE if there were any entries to replay.
1012  */
1013 boolean_t
1014 zil_destroy(zilog_t *zilog, boolean_t keep_first)
1015 {
1016 	const zil_header_t *zh = zilog->zl_header;
1017 	lwb_t *lwb;
1018 	dmu_tx_t *tx;
1019 	uint64_t txg;
1020 
1021 	/*
1022 	 * Wait for any previous destroy to complete.
1023 	 */
1024 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
1025 
1026 	zilog->zl_old_header = *zh;		/* debugging aid */
1027 
1028 	if (BP_IS_HOLE(&zh->zh_log))
1029 		return (B_FALSE);
1030 
1031 	tx = dmu_tx_create(zilog->zl_os);
1032 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT));
1033 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1034 	txg = dmu_tx_get_txg(tx);
1035 
1036 	mutex_enter(&zilog->zl_lock);
1037 
1038 	ASSERT3U(zilog->zl_destroy_txg, <, txg);
1039 	zilog->zl_destroy_txg = txg;
1040 	zilog->zl_keep_first = keep_first;
1041 
1042 	if (!list_is_empty(&zilog->zl_lwb_list)) {
1043 		ASSERT(zh->zh_claim_txg == 0);
1044 		VERIFY(!keep_first);
1045 		while ((lwb = list_remove_head(&zilog->zl_lwb_list)) != NULL) {
1046 			if (lwb->lwb_fastwrite)
1047 				metaslab_fastwrite_unmark(zilog->zl_spa,
1048 				    &lwb->lwb_blk);
1049 			if (lwb->lwb_buf != NULL)
1050 				zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1051 			zio_free(zilog->zl_spa, txg, &lwb->lwb_blk);
1052 			zil_free_lwb(zilog, lwb);
1053 		}
1054 	} else if (!keep_first) {
1055 		zil_destroy_sync(zilog, tx);
1056 	}
1057 	mutex_exit(&zilog->zl_lock);
1058 
1059 	dmu_tx_commit(tx);
1060 
1061 	return (B_TRUE);
1062 }
1063 
1064 void
1065 zil_destroy_sync(zilog_t *zilog, dmu_tx_t *tx)
1066 {
1067 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
1068 	(void) zil_parse(zilog, zil_free_log_block,
1069 	    zil_free_log_record, tx, zilog->zl_header->zh_claim_txg, B_FALSE);
1070 }
1071 
1072 int
1073 zil_claim(dsl_pool_t *dp, dsl_dataset_t *ds, void *txarg)
1074 {
1075 	dmu_tx_t *tx = txarg;
1076 	zilog_t *zilog;
1077 	uint64_t first_txg;
1078 	zil_header_t *zh;
1079 	objset_t *os;
1080 	int error;
1081 
1082 	error = dmu_objset_own_obj(dp, ds->ds_object,
1083 	    DMU_OST_ANY, B_FALSE, B_FALSE, FTAG, &os);
1084 	if (error != 0) {
1085 		/*
1086 		 * EBUSY indicates that the objset is inconsistent, in which
1087 		 * case it can not have a ZIL.
1088 		 */
1089 		if (error != EBUSY) {
1090 			cmn_err(CE_WARN, "can't open objset for %llu, error %u",
1091 			    (unsigned long long)ds->ds_object, error);
1092 		}
1093 
1094 		return (0);
1095 	}
1096 
1097 	zilog = dmu_objset_zil(os);
1098 	zh = zil_header_in_syncing_context(zilog);
1099 	ASSERT3U(tx->tx_txg, ==, spa_first_txg(zilog->zl_spa));
1100 	first_txg = spa_min_claim_txg(zilog->zl_spa);
1101 
1102 	/*
1103 	 * If the spa_log_state is not set to be cleared, check whether
1104 	 * the current uberblock is a checkpoint one and if the current
1105 	 * header has been claimed before moving on.
1106 	 *
1107 	 * If the current uberblock is a checkpointed uberblock then
1108 	 * one of the following scenarios took place:
1109 	 *
1110 	 * 1] We are currently rewinding to the checkpoint of the pool.
1111 	 * 2] We crashed in the middle of a checkpoint rewind but we
1112 	 *    did manage to write the checkpointed uberblock to the
1113 	 *    vdev labels, so when we tried to import the pool again
1114 	 *    the checkpointed uberblock was selected from the import
1115 	 *    procedure.
1116 	 *
1117 	 * In both cases we want to zero out all the ZIL blocks, except
1118 	 * the ones that have been claimed at the time of the checkpoint
1119 	 * (their zh_claim_txg != 0). The reason is that these blocks
1120 	 * may be corrupted since we may have reused their locations on
1121 	 * disk after we took the checkpoint.
1122 	 *
1123 	 * We could try to set spa_log_state to SPA_LOG_CLEAR earlier
1124 	 * when we first figure out whether the current uberblock is
1125 	 * checkpointed or not. Unfortunately, that would discard all
1126 	 * the logs, including the ones that are claimed, and we would
1127 	 * leak space.
1128 	 */
1129 	if (spa_get_log_state(zilog->zl_spa) == SPA_LOG_CLEAR ||
1130 	    (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1131 	    zh->zh_claim_txg == 0)) {
1132 		if (!BP_IS_HOLE(&zh->zh_log)) {
1133 			(void) zil_parse(zilog, zil_clear_log_block,
1134 			    zil_noop_log_record, tx, first_txg, B_FALSE);
1135 		}
1136 		BP_ZERO(&zh->zh_log);
1137 		if (os->os_encrypted)
1138 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1139 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
1140 		dmu_objset_disown(os, B_FALSE, FTAG);
1141 		return (0);
1142 	}
1143 
1144 	/*
1145 	 * If we are not rewinding and opening the pool normally, then
1146 	 * the min_claim_txg should be equal to the first txg of the pool.
1147 	 */
1148 	ASSERT3U(first_txg, ==, spa_first_txg(zilog->zl_spa));
1149 
1150 	/*
1151 	 * Claim all log blocks if we haven't already done so, and remember
1152 	 * the highest claimed sequence number.  This ensures that if we can
1153 	 * read only part of the log now (e.g. due to a missing device),
1154 	 * but we can read the entire log later, we will not try to replay
1155 	 * or destroy beyond the last block we successfully claimed.
1156 	 */
1157 	ASSERT3U(zh->zh_claim_txg, <=, first_txg);
1158 	if (zh->zh_claim_txg == 0 && !BP_IS_HOLE(&zh->zh_log)) {
1159 		(void) zil_parse(zilog, zil_claim_log_block,
1160 		    zil_claim_log_record, tx, first_txg, B_FALSE);
1161 		zh->zh_claim_txg = first_txg;
1162 		zh->zh_claim_blk_seq = zilog->zl_parse_blk_seq;
1163 		zh->zh_claim_lr_seq = zilog->zl_parse_lr_seq;
1164 		if (zilog->zl_parse_lr_count || zilog->zl_parse_blk_count > 1)
1165 			zh->zh_flags |= ZIL_REPLAY_NEEDED;
1166 		zh->zh_flags |= ZIL_CLAIM_LR_SEQ_VALID;
1167 		if (os->os_encrypted)
1168 			os->os_next_write_raw[tx->tx_txg & TXG_MASK] = B_TRUE;
1169 		dsl_dataset_dirty(dmu_objset_ds(os), tx);
1170 	}
1171 
1172 	ASSERT3U(first_txg, ==, (spa_last_synced_txg(zilog->zl_spa) + 1));
1173 	dmu_objset_disown(os, B_FALSE, FTAG);
1174 	return (0);
1175 }
1176 
1177 /*
1178  * Check the log by walking the log chain.
1179  * Checksum errors are ok as they indicate the end of the chain.
1180  * Any other error (no device or read failure) returns an error.
1181  */
1182 int
1183 zil_check_log_chain(dsl_pool_t *dp, dsl_dataset_t *ds, void *tx)
1184 {
1185 	(void) dp;
1186 	zilog_t *zilog;
1187 	objset_t *os;
1188 	blkptr_t *bp;
1189 	int error;
1190 
1191 	ASSERT(tx == NULL);
1192 
1193 	error = dmu_objset_from_ds(ds, &os);
1194 	if (error != 0) {
1195 		cmn_err(CE_WARN, "can't open objset %llu, error %d",
1196 		    (unsigned long long)ds->ds_object, error);
1197 		return (0);
1198 	}
1199 
1200 	zilog = dmu_objset_zil(os);
1201 	bp = (blkptr_t *)&zilog->zl_header->zh_log;
1202 
1203 	if (!BP_IS_HOLE(bp)) {
1204 		vdev_t *vd;
1205 		boolean_t valid = B_TRUE;
1206 
1207 		/*
1208 		 * Check the first block and determine if it's on a log device
1209 		 * which may have been removed or faulted prior to loading this
1210 		 * pool.  If so, there's no point in checking the rest of the
1211 		 * log as its content should have already been synced to the
1212 		 * pool.
1213 		 */
1214 		spa_config_enter(os->os_spa, SCL_STATE, FTAG, RW_READER);
1215 		vd = vdev_lookup_top(os->os_spa, DVA_GET_VDEV(&bp->blk_dva[0]));
1216 		if (vd->vdev_islog && vdev_is_dead(vd))
1217 			valid = vdev_log_state_valid(vd);
1218 		spa_config_exit(os->os_spa, SCL_STATE, FTAG);
1219 
1220 		if (!valid)
1221 			return (0);
1222 
1223 		/*
1224 		 * Check whether the current uberblock is checkpointed (e.g.
1225 		 * we are rewinding) and whether the current header has been
1226 		 * claimed or not. If it hasn't then skip verifying it. We
1227 		 * do this because its ZIL blocks may be part of the pool's
1228 		 * state before the rewind, which is no longer valid.
1229 		 */
1230 		zil_header_t *zh = zil_header_in_syncing_context(zilog);
1231 		if (zilog->zl_spa->spa_uberblock.ub_checkpoint_txg != 0 &&
1232 		    zh->zh_claim_txg == 0)
1233 			return (0);
1234 	}
1235 
1236 	/*
1237 	 * Because tx == NULL, zil_claim_log_block() will not actually claim
1238 	 * any blocks, but just determine whether it is possible to do so.
1239 	 * In addition to checking the log chain, zil_claim_log_block()
1240 	 * will invoke zio_claim() with a done func of spa_claim_notify(),
1241 	 * which will update spa_max_claim_txg.  See spa_load() for details.
1242 	 */
1243 	error = zil_parse(zilog, zil_claim_log_block, zil_claim_log_record, tx,
1244 	    zilog->zl_header->zh_claim_txg ? -1ULL :
1245 	    spa_min_claim_txg(os->os_spa), B_FALSE);
1246 
1247 	return ((error == ECKSUM || error == ENOENT) ? 0 : error);
1248 }
1249 
1250 /*
1251  * When an itx is "skipped", this function is used to properly mark the
1252  * waiter as "done, and signal any thread(s) waiting on it. An itx can
1253  * be skipped (and not committed to an lwb) for a variety of reasons,
1254  * one of them being that the itx was committed via spa_sync(), prior to
1255  * it being committed to an lwb; this can happen if a thread calling
1256  * zil_commit() is racing with spa_sync().
1257  */
1258 static void
1259 zil_commit_waiter_skip(zil_commit_waiter_t *zcw)
1260 {
1261 	mutex_enter(&zcw->zcw_lock);
1262 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1263 	zcw->zcw_done = B_TRUE;
1264 	cv_broadcast(&zcw->zcw_cv);
1265 	mutex_exit(&zcw->zcw_lock);
1266 }
1267 
1268 /*
1269  * This function is used when the given waiter is to be linked into an
1270  * lwb's "lwb_waiter" list; i.e. when the itx is committed to the lwb.
1271  * At this point, the waiter will no longer be referenced by the itx,
1272  * and instead, will be referenced by the lwb.
1273  */
1274 static void
1275 zil_commit_waiter_link_lwb(zil_commit_waiter_t *zcw, lwb_t *lwb)
1276 {
1277 	/*
1278 	 * The lwb_waiters field of the lwb is protected by the zilog's
1279 	 * zl_lock, thus it must be held when calling this function.
1280 	 */
1281 	ASSERT(MUTEX_HELD(&lwb->lwb_zilog->zl_lock));
1282 
1283 	mutex_enter(&zcw->zcw_lock);
1284 	ASSERT(!list_link_active(&zcw->zcw_node));
1285 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
1286 	ASSERT3P(lwb, !=, NULL);
1287 	ASSERT(lwb->lwb_state == LWB_STATE_OPENED ||
1288 	    lwb->lwb_state == LWB_STATE_ISSUED ||
1289 	    lwb->lwb_state == LWB_STATE_WRITE_DONE);
1290 
1291 	list_insert_tail(&lwb->lwb_waiters, zcw);
1292 	zcw->zcw_lwb = lwb;
1293 	mutex_exit(&zcw->zcw_lock);
1294 }
1295 
1296 /*
1297  * This function is used when zio_alloc_zil() fails to allocate a ZIL
1298  * block, and the given waiter must be linked to the "nolwb waiters"
1299  * list inside of zil_process_commit_list().
1300  */
1301 static void
1302 zil_commit_waiter_link_nolwb(zil_commit_waiter_t *zcw, list_t *nolwb)
1303 {
1304 	mutex_enter(&zcw->zcw_lock);
1305 	ASSERT(!list_link_active(&zcw->zcw_node));
1306 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
1307 	list_insert_tail(nolwb, zcw);
1308 	mutex_exit(&zcw->zcw_lock);
1309 }
1310 
1311 void
1312 zil_lwb_add_block(lwb_t *lwb, const blkptr_t *bp)
1313 {
1314 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1315 	avl_index_t where;
1316 	zil_vdev_node_t *zv, zvsearch;
1317 	int ndvas = BP_GET_NDVAS(bp);
1318 	int i;
1319 
1320 	if (zil_nocacheflush)
1321 		return;
1322 
1323 	mutex_enter(&lwb->lwb_vdev_lock);
1324 	for (i = 0; i < ndvas; i++) {
1325 		zvsearch.zv_vdev = DVA_GET_VDEV(&bp->blk_dva[i]);
1326 		if (avl_find(t, &zvsearch, &where) == NULL) {
1327 			zv = kmem_alloc(sizeof (*zv), KM_SLEEP);
1328 			zv->zv_vdev = zvsearch.zv_vdev;
1329 			avl_insert(t, zv, where);
1330 		}
1331 	}
1332 	mutex_exit(&lwb->lwb_vdev_lock);
1333 }
1334 
1335 static void
1336 zil_lwb_flush_defer(lwb_t *lwb, lwb_t *nlwb)
1337 {
1338 	avl_tree_t *src = &lwb->lwb_vdev_tree;
1339 	avl_tree_t *dst = &nlwb->lwb_vdev_tree;
1340 	void *cookie = NULL;
1341 	zil_vdev_node_t *zv;
1342 
1343 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1344 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
1345 	ASSERT3S(nlwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
1346 
1347 	/*
1348 	 * While 'lwb' is at a point in its lifetime where lwb_vdev_tree does
1349 	 * not need the protection of lwb_vdev_lock (it will only be modified
1350 	 * while holding zilog->zl_lock) as its writes and those of its
1351 	 * children have all completed.  The younger 'nlwb' may be waiting on
1352 	 * future writes to additional vdevs.
1353 	 */
1354 	mutex_enter(&nlwb->lwb_vdev_lock);
1355 	/*
1356 	 * Tear down the 'lwb' vdev tree, ensuring that entries which do not
1357 	 * exist in 'nlwb' are moved to it, freeing any would-be duplicates.
1358 	 */
1359 	while ((zv = avl_destroy_nodes(src, &cookie)) != NULL) {
1360 		avl_index_t where;
1361 
1362 		if (avl_find(dst, zv, &where) == NULL) {
1363 			avl_insert(dst, zv, where);
1364 		} else {
1365 			kmem_free(zv, sizeof (*zv));
1366 		}
1367 	}
1368 	mutex_exit(&nlwb->lwb_vdev_lock);
1369 }
1370 
1371 void
1372 zil_lwb_add_txg(lwb_t *lwb, uint64_t txg)
1373 {
1374 	lwb->lwb_max_txg = MAX(lwb->lwb_max_txg, txg);
1375 }
1376 
1377 /*
1378  * This function is a called after all vdevs associated with a given lwb
1379  * write have completed their DKIOCFLUSHWRITECACHE command; or as soon
1380  * as the lwb write completes, if "zil_nocacheflush" is set. Further,
1381  * all "previous" lwb's will have completed before this function is
1382  * called; i.e. this function is called for all previous lwbs before
1383  * it's called for "this" lwb (enforced via zio the dependencies
1384  * configured in zil_lwb_set_zio_dependency()).
1385  *
1386  * The intention is for this function to be called as soon as the
1387  * contents of an lwb are considered "stable" on disk, and will survive
1388  * any sudden loss of power. At this point, any threads waiting for the
1389  * lwb to reach this state are signalled, and the "waiter" structures
1390  * are marked "done".
1391  */
1392 static void
1393 zil_lwb_flush_vdevs_done(zio_t *zio)
1394 {
1395 	lwb_t *lwb = zio->io_private;
1396 	zilog_t *zilog = lwb->lwb_zilog;
1397 	zil_commit_waiter_t *zcw;
1398 	itx_t *itx;
1399 	uint64_t txg;
1400 	list_t itxs, waiters;
1401 
1402 	spa_config_exit(zilog->zl_spa, SCL_STATE, lwb);
1403 
1404 	list_create(&itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
1405 	list_create(&waiters, sizeof (zil_commit_waiter_t),
1406 	    offsetof(zil_commit_waiter_t, zcw_node));
1407 
1408 	hrtime_t t = gethrtime() - lwb->lwb_issued_timestamp;
1409 
1410 	mutex_enter(&zilog->zl_lock);
1411 
1412 	zilog->zl_last_lwb_latency = (zilog->zl_last_lwb_latency * 7 + t) / 8;
1413 
1414 	lwb->lwb_root_zio = NULL;
1415 
1416 	if (zilog->zl_last_lwb_opened == lwb) {
1417 		/*
1418 		 * Remember the highest committed log sequence number
1419 		 * for ztest. We only update this value when all the log
1420 		 * writes succeeded, because ztest wants to ASSERT that
1421 		 * it got the whole log chain.
1422 		 */
1423 		zilog->zl_commit_lr_seq = zilog->zl_lr_seq;
1424 	}
1425 
1426 	list_move_tail(&itxs, &lwb->lwb_itxs);
1427 	list_move_tail(&waiters, &lwb->lwb_waiters);
1428 	txg = lwb->lwb_issued_txg;
1429 
1430 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_WRITE_DONE);
1431 	lwb->lwb_state = LWB_STATE_FLUSH_DONE;
1432 
1433 	mutex_exit(&zilog->zl_lock);
1434 
1435 	while ((itx = list_remove_head(&itxs)) != NULL)
1436 		zil_itx_destroy(itx);
1437 	list_destroy(&itxs);
1438 
1439 	while ((zcw = list_remove_head(&waiters)) != NULL) {
1440 		mutex_enter(&zcw->zcw_lock);
1441 
1442 		zcw->zcw_lwb = NULL;
1443 		/*
1444 		 * We expect any ZIO errors from child ZIOs to have been
1445 		 * propagated "up" to this specific LWB's root ZIO, in
1446 		 * order for this error handling to work correctly. This
1447 		 * includes ZIO errors from either this LWB's write or
1448 		 * flush, as well as any errors from other dependent LWBs
1449 		 * (e.g. a root LWB ZIO that might be a child of this LWB).
1450 		 *
1451 		 * With that said, it's important to note that LWB flush
1452 		 * errors are not propagated up to the LWB root ZIO.
1453 		 * This is incorrect behavior, and results in VDEV flush
1454 		 * errors not being handled correctly here. See the
1455 		 * comment above the call to "zio_flush" for details.
1456 		 */
1457 
1458 		zcw->zcw_zio_error = zio->io_error;
1459 
1460 		ASSERT3B(zcw->zcw_done, ==, B_FALSE);
1461 		zcw->zcw_done = B_TRUE;
1462 		cv_broadcast(&zcw->zcw_cv);
1463 
1464 		mutex_exit(&zcw->zcw_lock);
1465 	}
1466 	list_destroy(&waiters);
1467 
1468 	mutex_enter(&zilog->zl_lwb_io_lock);
1469 	ASSERT3U(zilog->zl_lwb_inflight[txg & TXG_MASK], >, 0);
1470 	zilog->zl_lwb_inflight[txg & TXG_MASK]--;
1471 	if (zilog->zl_lwb_inflight[txg & TXG_MASK] == 0)
1472 		cv_broadcast(&zilog->zl_lwb_io_cv);
1473 	mutex_exit(&zilog->zl_lwb_io_lock);
1474 }
1475 
1476 /*
1477  * Wait for the completion of all issued write/flush of that txg provided.
1478  * It guarantees zil_lwb_flush_vdevs_done() is called and returned.
1479  */
1480 static void
1481 zil_lwb_flush_wait_all(zilog_t *zilog, uint64_t txg)
1482 {
1483 	ASSERT3U(txg, ==, spa_syncing_txg(zilog->zl_spa));
1484 
1485 	mutex_enter(&zilog->zl_lwb_io_lock);
1486 	while (zilog->zl_lwb_inflight[txg & TXG_MASK] > 0)
1487 		cv_wait(&zilog->zl_lwb_io_cv, &zilog->zl_lwb_io_lock);
1488 	mutex_exit(&zilog->zl_lwb_io_lock);
1489 
1490 #ifdef ZFS_DEBUG
1491 	mutex_enter(&zilog->zl_lock);
1492 	mutex_enter(&zilog->zl_lwb_io_lock);
1493 	lwb_t *lwb = list_head(&zilog->zl_lwb_list);
1494 	while (lwb != NULL && lwb->lwb_max_txg <= txg) {
1495 		if (lwb->lwb_issued_txg <= txg) {
1496 			ASSERT(lwb->lwb_state != LWB_STATE_ISSUED);
1497 			ASSERT(lwb->lwb_state != LWB_STATE_WRITE_DONE);
1498 			IMPLY(lwb->lwb_issued_txg > 0,
1499 			    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
1500 		}
1501 		IMPLY(lwb->lwb_state == LWB_STATE_WRITE_DONE ||
1502 		    lwb->lwb_state == LWB_STATE_FLUSH_DONE,
1503 		    lwb->lwb_buf == NULL);
1504 		lwb = list_next(&zilog->zl_lwb_list, lwb);
1505 	}
1506 	mutex_exit(&zilog->zl_lwb_io_lock);
1507 	mutex_exit(&zilog->zl_lock);
1508 #endif
1509 }
1510 
1511 /*
1512  * This is called when an lwb's write zio completes. The callback's
1513  * purpose is to issue the DKIOCFLUSHWRITECACHE commands for the vdevs
1514  * in the lwb's lwb_vdev_tree. The tree will contain the vdevs involved
1515  * in writing out this specific lwb's data, and in the case that cache
1516  * flushes have been deferred, vdevs involved in writing the data for
1517  * previous lwbs. The writes corresponding to all the vdevs in the
1518  * lwb_vdev_tree will have completed by the time this is called, due to
1519  * the zio dependencies configured in zil_lwb_set_zio_dependency(),
1520  * which takes deferred flushes into account. The lwb will be "done"
1521  * once zil_lwb_flush_vdevs_done() is called, which occurs in the zio
1522  * completion callback for the lwb's root zio.
1523  */
1524 static void
1525 zil_lwb_write_done(zio_t *zio)
1526 {
1527 	lwb_t *lwb = zio->io_private;
1528 	spa_t *spa = zio->io_spa;
1529 	zilog_t *zilog = lwb->lwb_zilog;
1530 	avl_tree_t *t = &lwb->lwb_vdev_tree;
1531 	void *cookie = NULL;
1532 	zil_vdev_node_t *zv;
1533 	lwb_t *nlwb;
1534 
1535 	ASSERT3S(spa_config_held(spa, SCL_STATE, RW_READER), !=, 0);
1536 
1537 	ASSERT(BP_GET_COMPRESS(zio->io_bp) == ZIO_COMPRESS_OFF);
1538 	ASSERT(BP_GET_TYPE(zio->io_bp) == DMU_OT_INTENT_LOG);
1539 	ASSERT(BP_GET_LEVEL(zio->io_bp) == 0);
1540 	ASSERT(BP_GET_BYTEORDER(zio->io_bp) == ZFS_HOST_BYTEORDER);
1541 	ASSERT(!BP_IS_GANG(zio->io_bp));
1542 	ASSERT(!BP_IS_HOLE(zio->io_bp));
1543 	ASSERT(BP_GET_FILL(zio->io_bp) == 0);
1544 
1545 	abd_free(zio->io_abd);
1546 	zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
1547 	lwb->lwb_buf = NULL;
1548 
1549 	mutex_enter(&zilog->zl_lock);
1550 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_ISSUED);
1551 	lwb->lwb_state = LWB_STATE_WRITE_DONE;
1552 	lwb->lwb_write_zio = NULL;
1553 	lwb->lwb_fastwrite = FALSE;
1554 	nlwb = list_next(&zilog->zl_lwb_list, lwb);
1555 	mutex_exit(&zilog->zl_lock);
1556 
1557 	if (avl_numnodes(t) == 0)
1558 		return;
1559 
1560 	/*
1561 	 * If there was an IO error, we're not going to call zio_flush()
1562 	 * on these vdevs, so we simply empty the tree and free the
1563 	 * nodes. We avoid calling zio_flush() since there isn't any
1564 	 * good reason for doing so, after the lwb block failed to be
1565 	 * written out.
1566 	 *
1567 	 * Additionally, we don't perform any further error handling at
1568 	 * this point (e.g. setting "zcw_zio_error" appropriately), as
1569 	 * we expect that to occur in "zil_lwb_flush_vdevs_done" (thus,
1570 	 * we expect any error seen here, to have been propagated to
1571 	 * that function).
1572 	 */
1573 	if (zio->io_error != 0) {
1574 		while ((zv = avl_destroy_nodes(t, &cookie)) != NULL)
1575 			kmem_free(zv, sizeof (*zv));
1576 		return;
1577 	}
1578 
1579 	/*
1580 	 * If this lwb does not have any threads waiting for it to
1581 	 * complete, we want to defer issuing the DKIOCFLUSHWRITECACHE
1582 	 * command to the vdevs written to by "this" lwb, and instead
1583 	 * rely on the "next" lwb to handle the DKIOCFLUSHWRITECACHE
1584 	 * command for those vdevs. Thus, we merge the vdev tree of
1585 	 * "this" lwb with the vdev tree of the "next" lwb in the list,
1586 	 * and assume the "next" lwb will handle flushing the vdevs (or
1587 	 * deferring the flush(s) again).
1588 	 *
1589 	 * This is a useful performance optimization, especially for
1590 	 * workloads with lots of async write activity and few sync
1591 	 * write and/or fsync activity, as it has the potential to
1592 	 * coalesce multiple flush commands to a vdev into one.
1593 	 */
1594 	if (list_is_empty(&lwb->lwb_waiters) && nlwb != NULL) {
1595 		zil_lwb_flush_defer(lwb, nlwb);
1596 		ASSERT(avl_is_empty(&lwb->lwb_vdev_tree));
1597 		return;
1598 	}
1599 
1600 	while ((zv = avl_destroy_nodes(t, &cookie)) != NULL) {
1601 		vdev_t *vd = vdev_lookup_top(spa, zv->zv_vdev);
1602 		if (vd != NULL && !vd->vdev_nowritecache) {
1603 			/*
1604 			 * The "ZIO_FLAG_DONT_PROPAGATE" is currently
1605 			 * always used within "zio_flush". This means,
1606 			 * any errors when flushing the vdev(s), will
1607 			 * (unfortunately) not be handled correctly,
1608 			 * since these "zio_flush" errors will not be
1609 			 * propagated up to "zil_lwb_flush_vdevs_done".
1610 			 */
1611 			zio_flush(lwb->lwb_root_zio, vd);
1612 		}
1613 		kmem_free(zv, sizeof (*zv));
1614 	}
1615 }
1616 
1617 static void
1618 zil_lwb_set_zio_dependency(zilog_t *zilog, lwb_t *lwb)
1619 {
1620 	lwb_t *last_lwb_opened = zilog->zl_last_lwb_opened;
1621 
1622 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1623 	ASSERT(MUTEX_HELD(&zilog->zl_lock));
1624 
1625 	/*
1626 	 * The zilog's "zl_last_lwb_opened" field is used to build the
1627 	 * lwb/zio dependency chain, which is used to preserve the
1628 	 * ordering of lwb completions that is required by the semantics
1629 	 * of the ZIL. Each new lwb zio becomes a parent of the
1630 	 * "previous" lwb zio, such that the new lwb's zio cannot
1631 	 * complete until the "previous" lwb's zio completes.
1632 	 *
1633 	 * This is required by the semantics of zil_commit(); the commit
1634 	 * waiters attached to the lwbs will be woken in the lwb zio's
1635 	 * completion callback, so this zio dependency graph ensures the
1636 	 * waiters are woken in the correct order (the same order the
1637 	 * lwbs were created).
1638 	 */
1639 	if (last_lwb_opened != NULL &&
1640 	    last_lwb_opened->lwb_state != LWB_STATE_FLUSH_DONE) {
1641 		ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1642 		    last_lwb_opened->lwb_state == LWB_STATE_ISSUED ||
1643 		    last_lwb_opened->lwb_state == LWB_STATE_WRITE_DONE);
1644 
1645 		ASSERT3P(last_lwb_opened->lwb_root_zio, !=, NULL);
1646 		zio_add_child(lwb->lwb_root_zio,
1647 		    last_lwb_opened->lwb_root_zio);
1648 
1649 		/*
1650 		 * If the previous lwb's write hasn't already completed,
1651 		 * we also want to order the completion of the lwb write
1652 		 * zios (above, we only order the completion of the lwb
1653 		 * root zios). This is required because of how we can
1654 		 * defer the DKIOCFLUSHWRITECACHE commands for each lwb.
1655 		 *
1656 		 * When the DKIOCFLUSHWRITECACHE commands are deferred,
1657 		 * the previous lwb will rely on this lwb to flush the
1658 		 * vdevs written to by that previous lwb. Thus, we need
1659 		 * to ensure this lwb doesn't issue the flush until
1660 		 * after the previous lwb's write completes. We ensure
1661 		 * this ordering by setting the zio parent/child
1662 		 * relationship here.
1663 		 *
1664 		 * Without this relationship on the lwb's write zio,
1665 		 * it's possible for this lwb's write to complete prior
1666 		 * to the previous lwb's write completing; and thus, the
1667 		 * vdevs for the previous lwb would be flushed prior to
1668 		 * that lwb's data being written to those vdevs (the
1669 		 * vdevs are flushed in the lwb write zio's completion
1670 		 * handler, zil_lwb_write_done()).
1671 		 */
1672 		if (last_lwb_opened->lwb_state != LWB_STATE_WRITE_DONE) {
1673 			ASSERT(last_lwb_opened->lwb_state == LWB_STATE_OPENED ||
1674 			    last_lwb_opened->lwb_state == LWB_STATE_ISSUED);
1675 
1676 			ASSERT3P(last_lwb_opened->lwb_write_zio, !=, NULL);
1677 			zio_add_child(lwb->lwb_write_zio,
1678 			    last_lwb_opened->lwb_write_zio);
1679 		}
1680 	}
1681 }
1682 
1683 
1684 /*
1685  * This function's purpose is to "open" an lwb such that it is ready to
1686  * accept new itxs being committed to it. To do this, the lwb's zio
1687  * structures are created, and linked to the lwb. This function is
1688  * idempotent; if the passed in lwb has already been opened, this
1689  * function is essentially a no-op.
1690  */
1691 static void
1692 zil_lwb_write_open(zilog_t *zilog, lwb_t *lwb)
1693 {
1694 	zbookmark_phys_t zb;
1695 	zio_priority_t prio;
1696 
1697 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1698 	ASSERT3P(lwb, !=, NULL);
1699 	EQUIV(lwb->lwb_root_zio == NULL, lwb->lwb_state == LWB_STATE_CLOSED);
1700 	EQUIV(lwb->lwb_root_zio != NULL, lwb->lwb_state == LWB_STATE_OPENED);
1701 
1702 	if (lwb->lwb_root_zio != NULL)
1703 		return;
1704 
1705 	lwb->lwb_root_zio = zio_root(zilog->zl_spa,
1706 	    zil_lwb_flush_vdevs_done, lwb, ZIO_FLAG_CANFAIL);
1707 
1708 	abd_t *lwb_abd = abd_get_from_buf(lwb->lwb_buf,
1709 	    BP_GET_LSIZE(&lwb->lwb_blk));
1710 
1711 	if (!lwb->lwb_slog || zilog->zl_cur_used <= zil_slog_bulk)
1712 		prio = ZIO_PRIORITY_SYNC_WRITE;
1713 	else
1714 		prio = ZIO_PRIORITY_ASYNC_WRITE;
1715 
1716 	SET_BOOKMARK(&zb, lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_OBJSET],
1717 	    ZB_ZIL_OBJECT, ZB_ZIL_LEVEL,
1718 	    lwb->lwb_blk.blk_cksum.zc_word[ZIL_ZC_SEQ]);
1719 
1720 	/* Lock so zil_sync() doesn't fastwrite_unmark after zio is created */
1721 	mutex_enter(&zilog->zl_lock);
1722 	if (!lwb->lwb_fastwrite) {
1723 		metaslab_fastwrite_mark(zilog->zl_spa, &lwb->lwb_blk);
1724 		lwb->lwb_fastwrite = 1;
1725 	}
1726 
1727 	lwb->lwb_write_zio = zio_rewrite(lwb->lwb_root_zio, zilog->zl_spa, 0,
1728 	    &lwb->lwb_blk, lwb_abd, BP_GET_LSIZE(&lwb->lwb_blk),
1729 	    zil_lwb_write_done, lwb, prio,
1730 	    ZIO_FLAG_CANFAIL | ZIO_FLAG_FASTWRITE, &zb);
1731 
1732 	lwb->lwb_state = LWB_STATE_OPENED;
1733 
1734 	zil_lwb_set_zio_dependency(zilog, lwb);
1735 	zilog->zl_last_lwb_opened = lwb;
1736 	mutex_exit(&zilog->zl_lock);
1737 }
1738 
1739 /*
1740  * Define a limited set of intent log block sizes.
1741  *
1742  * These must be a multiple of 4KB. Note only the amount used (again
1743  * aligned to 4KB) actually gets written. However, we can't always just
1744  * allocate SPA_OLD_MAXBLOCKSIZE as the slog space could be exhausted.
1745  */
1746 static const struct {
1747 	uint64_t	limit;
1748 	uint64_t	blksz;
1749 } zil_block_buckets[] = {
1750 	{ 4096,		4096 },			/* non TX_WRITE */
1751 	{ 8192 + 4096,	8192 + 4096 },		/* database */
1752 	{ 32768 + 4096,	32768 + 4096 },		/* NFS writes */
1753 	{ 65536 + 4096,	65536 + 4096 },		/* 64KB writes */
1754 	{ 131072,	131072 },		/* < 128KB writes */
1755 	{ 131072 +4096,	65536 + 4096 },		/* 128KB writes */
1756 	{ UINT64_MAX,	SPA_OLD_MAXBLOCKSIZE},	/* > 128KB writes */
1757 };
1758 
1759 /*
1760  * Maximum block size used by the ZIL.  This is picked up when the ZIL is
1761  * initialized.  Otherwise this should not be used directly; see
1762  * zl_max_block_size instead.
1763  */
1764 static uint_t zil_maxblocksize = SPA_OLD_MAXBLOCKSIZE;
1765 
1766 /*
1767  * Close the log block for being issued and allocate the next one.
1768  * Has to be called under zl_issuer_lock to chain more lwbs.
1769  */
1770 static lwb_t *
1771 zil_lwb_write_close(zilog_t *zilog, lwb_t *lwb)
1772 {
1773 	lwb_t *nlwb = NULL;
1774 	zil_chain_t *zilc;
1775 	spa_t *spa = zilog->zl_spa;
1776 	blkptr_t *bp;
1777 	dmu_tx_t *tx;
1778 	uint64_t txg;
1779 	uint64_t zil_blksz;
1780 	int i, error;
1781 	boolean_t slog;
1782 
1783 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1784 	ASSERT3P(lwb->lwb_root_zio, !=, NULL);
1785 	ASSERT3P(lwb->lwb_write_zio, !=, NULL);
1786 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
1787 
1788 	/*
1789 	 * If this lwb includes indirect writes, we have to commit before
1790 	 * creating the transaction, otherwise we may end up in dead lock.
1791 	 */
1792 	if (lwb->lwb_indirect) {
1793 		for (itx_t *itx = list_head(&lwb->lwb_itxs); itx;
1794 		    itx = list_next(&lwb->lwb_itxs, itx))
1795 			zil_lwb_commit(zilog, lwb, itx);
1796 		lwb->lwb_nused = lwb->lwb_nfilled;
1797 	}
1798 
1799 	/*
1800 	 * Allocate the next block and save its address in this block
1801 	 * before writing it in order to establish the log chain.
1802 	 */
1803 
1804 	tx = dmu_tx_create(zilog->zl_os);
1805 
1806 	/*
1807 	 * Since we are not going to create any new dirty data, and we
1808 	 * can even help with clearing the existing dirty data, we
1809 	 * should not be subject to the dirty data based delays. We
1810 	 * use TXG_NOTHROTTLE to bypass the delay mechanism.
1811 	 */
1812 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
1813 
1814 	dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
1815 	txg = dmu_tx_get_txg(tx);
1816 
1817 	mutex_enter(&zilog->zl_lwb_io_lock);
1818 	lwb->lwb_issued_txg = txg;
1819 	zilog->zl_lwb_inflight[txg & TXG_MASK]++;
1820 	zilog->zl_lwb_max_issued_txg = MAX(txg, zilog->zl_lwb_max_issued_txg);
1821 	mutex_exit(&zilog->zl_lwb_io_lock);
1822 
1823 	/*
1824 	 * Log blocks are pre-allocated. Here we select the size of the next
1825 	 * block, based on size used in the last block.
1826 	 * - first find the smallest bucket that will fit the block from a
1827 	 *   limited set of block sizes. This is because it's faster to write
1828 	 *   blocks allocated from the same metaslab as they are adjacent or
1829 	 *   close.
1830 	 * - next find the maximum from the new suggested size and an array of
1831 	 *   previous sizes. This lessens a picket fence effect of wrongly
1832 	 *   guessing the size if we have a stream of say 2k, 64k, 2k, 64k
1833 	 *   requests.
1834 	 *
1835 	 * Note we only write what is used, but we can't just allocate
1836 	 * the maximum block size because we can exhaust the available
1837 	 * pool log space.
1838 	 */
1839 	zil_blksz = zilog->zl_cur_used + sizeof (zil_chain_t);
1840 	for (i = 0; zil_blksz > zil_block_buckets[i].limit; i++)
1841 		continue;
1842 	zil_blksz = MIN(zil_block_buckets[i].blksz, zilog->zl_max_block_size);
1843 	zilog->zl_prev_blks[zilog->zl_prev_rotor] = zil_blksz;
1844 	for (i = 0; i < ZIL_PREV_BLKS; i++)
1845 		zil_blksz = MAX(zil_blksz, zilog->zl_prev_blks[i]);
1846 	DTRACE_PROBE3(zil__block__size, zilog_t *, zilog,
1847 	    uint64_t, zil_blksz,
1848 	    uint64_t, zilog->zl_prev_blks[zilog->zl_prev_rotor]);
1849 	zilog->zl_prev_rotor = (zilog->zl_prev_rotor + 1) & (ZIL_PREV_BLKS - 1);
1850 
1851 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2)
1852 		zilc = (zil_chain_t *)lwb->lwb_buf;
1853 	else
1854 		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1855 	bp = &zilc->zc_next_blk;
1856 	BP_ZERO(bp);
1857 	error = zio_alloc_zil(spa, zilog->zl_os, txg, bp, zil_blksz, &slog);
1858 	if (error == 0) {
1859 		ASSERT3U(bp->blk_birth, ==, txg);
1860 		bp->blk_cksum = lwb->lwb_blk.blk_cksum;
1861 		bp->blk_cksum.zc_word[ZIL_ZC_SEQ]++;
1862 
1863 		/*
1864 		 * Allocate a new log write block (lwb).
1865 		 */
1866 		nlwb = zil_alloc_lwb(zilog, bp, slog, txg, TRUE);
1867 	}
1868 
1869 	lwb->lwb_state = LWB_STATE_ISSUED;
1870 
1871 	dmu_tx_commit(tx);
1872 
1873 	/*
1874 	 * If there was an allocation failure then nlwb will be null which
1875 	 * forces a txg_wait_synced().
1876 	 */
1877 	return (nlwb);
1878 }
1879 
1880 /*
1881  * Finalize previously closed block and issue the write zio.
1882  * Does not require locking.
1883  */
1884 static void
1885 zil_lwb_write_issue(zilog_t *zilog, lwb_t *lwb)
1886 {
1887 	zil_chain_t *zilc;
1888 	int wsz;
1889 
1890 	/* Actually fill the lwb with the data if not yet. */
1891 	if (!lwb->lwb_indirect) {
1892 		for (itx_t *itx = list_head(&lwb->lwb_itxs); itx;
1893 		    itx = list_next(&lwb->lwb_itxs, itx))
1894 			zil_lwb_commit(zilog, lwb, itx);
1895 		lwb->lwb_nused = lwb->lwb_nfilled;
1896 	}
1897 
1898 	if (BP_GET_CHECKSUM(&lwb->lwb_blk) == ZIO_CHECKSUM_ZILOG2) {
1899 		/* For Slim ZIL only write what is used. */
1900 		wsz = P2ROUNDUP_TYPED(lwb->lwb_nused, ZIL_MIN_BLKSZ, int);
1901 		ASSERT3S(wsz, <=, lwb->lwb_sz);
1902 		zio_shrink(lwb->lwb_write_zio, wsz);
1903 		wsz = lwb->lwb_write_zio->io_size;
1904 
1905 		zilc = (zil_chain_t *)lwb->lwb_buf;
1906 	} else {
1907 		wsz = lwb->lwb_sz;
1908 		zilc = (zil_chain_t *)(lwb->lwb_buf + lwb->lwb_sz);
1909 	}
1910 	zilc->zc_pad = 0;
1911 	zilc->zc_nused = lwb->lwb_nused;
1912 	zilc->zc_eck.zec_cksum = lwb->lwb_blk.blk_cksum;
1913 
1914 	/*
1915 	 * clear unused data for security
1916 	 */
1917 	memset(lwb->lwb_buf + lwb->lwb_nused, 0, wsz - lwb->lwb_nused);
1918 
1919 	if (lwb->lwb_slog) {
1920 		ZIL_STAT_BUMP(zilog, zil_itx_metaslab_slog_count);
1921 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_bytes,
1922 		    lwb->lwb_nused);
1923 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_write,
1924 		    wsz);
1925 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_slog_alloc,
1926 		    BP_GET_LSIZE(&lwb->lwb_blk));
1927 	} else {
1928 		ZIL_STAT_BUMP(zilog, zil_itx_metaslab_normal_count);
1929 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_bytes,
1930 		    lwb->lwb_nused);
1931 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_write,
1932 		    wsz);
1933 		ZIL_STAT_INCR(zilog, zil_itx_metaslab_normal_alloc,
1934 		    BP_GET_LSIZE(&lwb->lwb_blk));
1935 	}
1936 	spa_config_enter(zilog->zl_spa, SCL_STATE, lwb, RW_READER);
1937 	zil_lwb_add_block(lwb, &lwb->lwb_blk);
1938 	lwb->lwb_issued_timestamp = gethrtime();
1939 	zio_nowait(lwb->lwb_root_zio);
1940 	zio_nowait(lwb->lwb_write_zio);
1941 }
1942 
1943 /*
1944  * Maximum amount of data that can be put into single log block.
1945  */
1946 uint64_t
1947 zil_max_log_data(zilog_t *zilog, size_t hdrsize)
1948 {
1949 	return (zilog->zl_max_block_size - sizeof (zil_chain_t) - hdrsize);
1950 }
1951 
1952 /*
1953  * Maximum amount of log space we agree to waste to reduce number of
1954  * WR_NEED_COPY chunks to reduce zl_get_data() overhead (~12%).
1955  */
1956 static inline uint64_t
1957 zil_max_waste_space(zilog_t *zilog)
1958 {
1959 	return (zil_max_log_data(zilog, sizeof (lr_write_t)) / 8);
1960 }
1961 
1962 /*
1963  * Maximum amount of write data for WR_COPIED.  For correctness, consumers
1964  * must fall back to WR_NEED_COPY if we can't fit the entire record into one
1965  * maximum sized log block, because each WR_COPIED record must fit in a
1966  * single log block.  For space efficiency, we want to fit two records into a
1967  * max-sized log block.
1968  */
1969 uint64_t
1970 zil_max_copied_data(zilog_t *zilog)
1971 {
1972 	return ((zilog->zl_max_block_size - sizeof (zil_chain_t)) / 2 -
1973 	    sizeof (lr_write_t));
1974 }
1975 
1976 /*
1977  * Estimate space needed in the lwb for the itx.  Allocate more lwbs or
1978  * split the itx as needed, but don't touch the actual transaction data.
1979  * Has to be called under zl_issuer_lock to call zil_lwb_write_close()
1980  * to chain more lwbs.
1981  */
1982 static lwb_t *
1983 zil_lwb_assign(zilog_t *zilog, lwb_t *lwb, itx_t *itx, list_t *ilwbs)
1984 {
1985 	itx_t *citx;
1986 	lr_t *lr, *clr;
1987 	lr_write_t *lrw;
1988 	uint64_t dlen, dnow, lwb_sp, reclen, max_log_data;
1989 
1990 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
1991 	ASSERT3P(lwb, !=, NULL);
1992 	ASSERT3P(lwb->lwb_buf, !=, NULL);
1993 
1994 	zil_lwb_write_open(zilog, lwb);
1995 
1996 	lr = &itx->itx_lr;
1997 	lrw = (lr_write_t *)lr;
1998 
1999 	/*
2000 	 * A commit itx doesn't represent any on-disk state; instead
2001 	 * it's simply used as a place holder on the commit list, and
2002 	 * provides a mechanism for attaching a "commit waiter" onto the
2003 	 * correct lwb (such that the waiter can be signalled upon
2004 	 * completion of that lwb). Thus, we don't process this itx's
2005 	 * log record if it's a commit itx (these itx's don't have log
2006 	 * records), and instead link the itx's waiter onto the lwb's
2007 	 * list of waiters.
2008 	 *
2009 	 * For more details, see the comment above zil_commit().
2010 	 */
2011 	if (lr->lrc_txtype == TX_COMMIT) {
2012 		mutex_enter(&zilog->zl_lock);
2013 		zil_commit_waiter_link_lwb(itx->itx_private, lwb);
2014 		itx->itx_private = NULL;
2015 		mutex_exit(&zilog->zl_lock);
2016 		list_insert_tail(&lwb->lwb_itxs, itx);
2017 		return (lwb);
2018 	}
2019 
2020 	if (lr->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
2021 		dlen = P2ROUNDUP_TYPED(
2022 		    lrw->lr_length, sizeof (uint64_t), uint64_t);
2023 	} else {
2024 		dlen = 0;
2025 	}
2026 	reclen = lr->lrc_reclen;
2027 	zilog->zl_cur_used += (reclen + dlen);
2028 
2029 cont:
2030 	/*
2031 	 * If this record won't fit in the current log block, start a new one.
2032 	 * For WR_NEED_COPY optimize layout for minimal number of chunks.
2033 	 */
2034 	lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
2035 	max_log_data = zil_max_log_data(zilog, sizeof (lr_write_t));
2036 	if (reclen > lwb_sp || (reclen + dlen > lwb_sp &&
2037 	    lwb_sp < zil_max_waste_space(zilog) &&
2038 	    (dlen % max_log_data == 0 ||
2039 	    lwb_sp < reclen + dlen % max_log_data))) {
2040 		list_insert_tail(ilwbs, lwb);
2041 		lwb = zil_lwb_write_close(zilog, lwb);
2042 		if (lwb == NULL)
2043 			return (NULL);
2044 		zil_lwb_write_open(zilog, lwb);
2045 		lwb_sp = lwb->lwb_sz - lwb->lwb_nused;
2046 
2047 		/*
2048 		 * There must be enough space in the new, empty log block to
2049 		 * hold reclen.  For WR_COPIED, we need to fit the whole
2050 		 * record in one block, and reclen is the header size + the
2051 		 * data size. For WR_NEED_COPY, we can create multiple
2052 		 * records, splitting the data into multiple blocks, so we
2053 		 * only need to fit one word of data per block; in this case
2054 		 * reclen is just the header size (no data).
2055 		 */
2056 		ASSERT3U(reclen + MIN(dlen, sizeof (uint64_t)), <=, lwb_sp);
2057 	}
2058 
2059 	dnow = MIN(dlen, lwb_sp - reclen);
2060 	if (dlen > dnow) {
2061 		ASSERT3U(lr->lrc_txtype, ==, TX_WRITE);
2062 		ASSERT3U(itx->itx_wr_state, ==, WR_NEED_COPY);
2063 		citx = zil_itx_clone(itx);
2064 		clr = &citx->itx_lr;
2065 		lr_write_t *clrw = (lr_write_t *)clr;
2066 		clrw->lr_length = dnow;
2067 		lrw->lr_offset += dnow;
2068 		lrw->lr_length -= dnow;
2069 	} else {
2070 		citx = itx;
2071 		clr = lr;
2072 	}
2073 
2074 	/*
2075 	 * We're actually making an entry, so update lrc_seq to be the
2076 	 * log record sequence number.  Note that this is generally not
2077 	 * equal to the itx sequence number because not all transactions
2078 	 * are synchronous, and sometimes spa_sync() gets there first.
2079 	 */
2080 	clr->lrc_seq = ++zilog->zl_lr_seq;
2081 
2082 	lwb->lwb_nused += reclen + dnow;
2083 	ASSERT3U(lwb->lwb_nused, <=, lwb->lwb_sz);
2084 	ASSERT0(P2PHASE(lwb->lwb_nused, sizeof (uint64_t)));
2085 
2086 	zil_lwb_add_txg(lwb, lr->lrc_txg);
2087 	list_insert_tail(&lwb->lwb_itxs, citx);
2088 
2089 	dlen -= dnow;
2090 	if (dlen > 0) {
2091 		zilog->zl_cur_used += reclen;
2092 		goto cont;
2093 	}
2094 
2095 	/*
2096 	 * We have to really issue all queued LWBs before we may have to
2097 	 * wait for a txg sync.  Otherwise we may end up in a dead lock.
2098 	 */
2099 	if (lr->lrc_txtype == TX_WRITE) {
2100 		boolean_t frozen = lr->lrc_txg > spa_freeze_txg(zilog->zl_spa);
2101 		if (frozen || itx->itx_wr_state == WR_INDIRECT) {
2102 			lwb_t *tlwb;
2103 			while ((tlwb = list_remove_head(ilwbs)) != NULL)
2104 				zil_lwb_write_issue(zilog, tlwb);
2105 		}
2106 		if (itx->itx_wr_state == WR_INDIRECT)
2107 			lwb->lwb_indirect = B_TRUE;
2108 		if (frozen)
2109 			txg_wait_synced(zilog->zl_dmu_pool, lr->lrc_txg);
2110 	}
2111 
2112 	return (lwb);
2113 }
2114 
2115 /*
2116  * Fill the actual transaction data into the lwb, following zil_lwb_assign().
2117  * Does not require locking.
2118  */
2119 static void
2120 zil_lwb_commit(zilog_t *zilog, lwb_t *lwb, itx_t *itx)
2121 {
2122 	lr_t *lr, *lrb;
2123 	lr_write_t *lrw, *lrwb;
2124 	char *lr_buf;
2125 	uint64_t dlen, reclen;
2126 
2127 	lr = &itx->itx_lr;
2128 	lrw = (lr_write_t *)lr;
2129 
2130 	if (lr->lrc_txtype == TX_COMMIT)
2131 		return;
2132 
2133 	if (lr->lrc_txtype == TX_WRITE && itx->itx_wr_state == WR_NEED_COPY) {
2134 		dlen = P2ROUNDUP_TYPED(
2135 		    lrw->lr_length, sizeof (uint64_t), uint64_t);
2136 	} else {
2137 		dlen = 0;
2138 	}
2139 	reclen = lr->lrc_reclen;
2140 	ASSERT3U(reclen + dlen, <=, lwb->lwb_nused - lwb->lwb_nfilled);
2141 
2142 	lr_buf = lwb->lwb_buf + lwb->lwb_nfilled;
2143 	memcpy(lr_buf, lr, reclen);
2144 	lrb = (lr_t *)lr_buf;		/* Like lr, but inside lwb. */
2145 	lrwb = (lr_write_t *)lrb;	/* Like lrw, but inside lwb. */
2146 
2147 	ZIL_STAT_BUMP(zilog, zil_itx_count);
2148 
2149 	/*
2150 	 * If it's a write, fetch the data or get its blkptr as appropriate.
2151 	 */
2152 	if (lr->lrc_txtype == TX_WRITE) {
2153 		if (itx->itx_wr_state == WR_COPIED) {
2154 			ZIL_STAT_BUMP(zilog, zil_itx_copied_count);
2155 			ZIL_STAT_INCR(zilog, zil_itx_copied_bytes,
2156 			    lrw->lr_length);
2157 		} else {
2158 			char *dbuf;
2159 			int error;
2160 
2161 			if (itx->itx_wr_state == WR_NEED_COPY) {
2162 				dbuf = lr_buf + reclen;
2163 				lrb->lrc_reclen += dlen;
2164 				ZIL_STAT_BUMP(zilog, zil_itx_needcopy_count);
2165 				ZIL_STAT_INCR(zilog, zil_itx_needcopy_bytes,
2166 				    dlen);
2167 			} else {
2168 				ASSERT3S(itx->itx_wr_state, ==, WR_INDIRECT);
2169 				dbuf = NULL;
2170 				ZIL_STAT_BUMP(zilog, zil_itx_indirect_count);
2171 				ZIL_STAT_INCR(zilog, zil_itx_indirect_bytes,
2172 				    lrw->lr_length);
2173 			}
2174 
2175 			/*
2176 			 * We pass in the "lwb_write_zio" rather than
2177 			 * "lwb_root_zio" so that the "lwb_write_zio"
2178 			 * becomes the parent of any zio's created by
2179 			 * the "zl_get_data" callback. The vdevs are
2180 			 * flushed after the "lwb_write_zio" completes,
2181 			 * so we want to make sure that completion
2182 			 * callback waits for these additional zio's,
2183 			 * such that the vdevs used by those zio's will
2184 			 * be included in the lwb's vdev tree, and those
2185 			 * vdevs will be properly flushed. If we passed
2186 			 * in "lwb_root_zio" here, then these additional
2187 			 * vdevs may not be flushed; e.g. if these zio's
2188 			 * completed after "lwb_write_zio" completed.
2189 			 */
2190 			error = zilog->zl_get_data(itx->itx_private,
2191 			    itx->itx_gen, lrwb, dbuf, lwb,
2192 			    lwb->lwb_write_zio);
2193 			if (dbuf != NULL && error == 0) {
2194 				/* Zero any padding bytes in the last block. */
2195 				memset((char *)dbuf + lrwb->lr_length, 0,
2196 				    dlen - lrwb->lr_length);
2197 			}
2198 
2199 			/*
2200 			 * Typically, the only return values we should see from
2201 			 * ->zl_get_data() are 0, EIO, ENOENT, EEXIST or
2202 			 *  EALREADY. However, it is also possible to see other
2203 			 *  error values such as ENOSPC or EINVAL from
2204 			 *  dmu_read() -> dnode_hold() -> dnode_hold_impl() or
2205 			 *  ENXIO as well as a multitude of others from the
2206 			 *  block layer through dmu_buf_hold() -> dbuf_read()
2207 			 *  -> zio_wait(), as well as through dmu_read() ->
2208 			 *  dnode_hold() -> dnode_hold_impl() -> dbuf_read() ->
2209 			 *  zio_wait(). When these errors happen, we can assume
2210 			 *  that neither an immediate write nor an indirect
2211 			 *  write occurred, so we need to fall back to
2212 			 *  txg_wait_synced(). This is unusual, so we print to
2213 			 *  dmesg whenever one of these errors occurs.
2214 			 */
2215 			switch (error) {
2216 			case 0:
2217 				break;
2218 			default:
2219 				cmn_err(CE_WARN, "zil_lwb_commit() received "
2220 				    "unexpected error %d from ->zl_get_data()"
2221 				    ". Falling back to txg_wait_synced().",
2222 				    error);
2223 				zfs_fallthrough;
2224 			case EIO:
2225 				if (lwb->lwb_indirect) {
2226 					txg_wait_synced(zilog->zl_dmu_pool,
2227 					    lr->lrc_txg);
2228 				} else {
2229 					lwb->lwb_write_zio->io_error = error;
2230 				}
2231 				zfs_fallthrough;
2232 			case ENOENT:
2233 				zfs_fallthrough;
2234 			case EEXIST:
2235 				zfs_fallthrough;
2236 			case EALREADY:
2237 				return;
2238 			}
2239 		}
2240 	}
2241 
2242 	lwb->lwb_nfilled += reclen + dlen;
2243 	ASSERT3S(lwb->lwb_nfilled, <=, lwb->lwb_nused);
2244 	ASSERT0(P2PHASE(lwb->lwb_nfilled, sizeof (uint64_t)));
2245 }
2246 
2247 itx_t *
2248 zil_itx_create(uint64_t txtype, size_t olrsize)
2249 {
2250 	size_t itxsize, lrsize;
2251 	itx_t *itx;
2252 
2253 	lrsize = P2ROUNDUP_TYPED(olrsize, sizeof (uint64_t), size_t);
2254 	itxsize = offsetof(itx_t, itx_lr) + lrsize;
2255 
2256 	itx = zio_data_buf_alloc(itxsize);
2257 	itx->itx_lr.lrc_txtype = txtype;
2258 	itx->itx_lr.lrc_reclen = lrsize;
2259 	itx->itx_lr.lrc_seq = 0;	/* defensive */
2260 	memset((char *)&itx->itx_lr + olrsize, 0, lrsize - olrsize);
2261 	itx->itx_sync = B_TRUE;		/* default is synchronous */
2262 	itx->itx_callback = NULL;
2263 	itx->itx_callback_data = NULL;
2264 	itx->itx_size = itxsize;
2265 
2266 	return (itx);
2267 }
2268 
2269 static itx_t *
2270 zil_itx_clone(itx_t *oitx)
2271 {
2272 	itx_t *itx = zio_data_buf_alloc(oitx->itx_size);
2273 	memcpy(itx, oitx, oitx->itx_size);
2274 	itx->itx_callback = NULL;
2275 	itx->itx_callback_data = NULL;
2276 	return (itx);
2277 }
2278 
2279 void
2280 zil_itx_destroy(itx_t *itx)
2281 {
2282 	IMPLY(itx->itx_lr.lrc_txtype == TX_COMMIT, itx->itx_callback == NULL);
2283 	IMPLY(itx->itx_callback != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2284 
2285 	if (itx->itx_callback != NULL)
2286 		itx->itx_callback(itx->itx_callback_data);
2287 
2288 	zio_data_buf_free(itx, itx->itx_size);
2289 }
2290 
2291 /*
2292  * Free up the sync and async itxs. The itxs_t has already been detached
2293  * so no locks are needed.
2294  */
2295 static void
2296 zil_itxg_clean(void *arg)
2297 {
2298 	itx_t *itx;
2299 	list_t *list;
2300 	avl_tree_t *t;
2301 	void *cookie;
2302 	itxs_t *itxs = arg;
2303 	itx_async_node_t *ian;
2304 
2305 	list = &itxs->i_sync_list;
2306 	while ((itx = list_remove_head(list)) != NULL) {
2307 		/*
2308 		 * In the general case, commit itxs will not be found
2309 		 * here, as they'll be committed to an lwb via
2310 		 * zil_lwb_assign(), and free'd in that function. Having
2311 		 * said that, it is still possible for commit itxs to be
2312 		 * found here, due to the following race:
2313 		 *
2314 		 *	- a thread calls zil_commit() which assigns the
2315 		 *	  commit itx to a per-txg i_sync_list
2316 		 *	- zil_itxg_clean() is called (e.g. via spa_sync())
2317 		 *	  while the waiter is still on the i_sync_list
2318 		 *
2319 		 * There's nothing to prevent syncing the txg while the
2320 		 * waiter is on the i_sync_list. This normally doesn't
2321 		 * happen because spa_sync() is slower than zil_commit(),
2322 		 * but if zil_commit() calls txg_wait_synced() (e.g.
2323 		 * because zil_create() or zil_commit_writer_stall() is
2324 		 * called) we will hit this case.
2325 		 */
2326 		if (itx->itx_lr.lrc_txtype == TX_COMMIT)
2327 			zil_commit_waiter_skip(itx->itx_private);
2328 
2329 		zil_itx_destroy(itx);
2330 	}
2331 
2332 	cookie = NULL;
2333 	t = &itxs->i_async_tree;
2334 	while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2335 		list = &ian->ia_list;
2336 		while ((itx = list_remove_head(list)) != NULL) {
2337 			/* commit itxs should never be on the async lists. */
2338 			ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2339 			zil_itx_destroy(itx);
2340 		}
2341 		list_destroy(list);
2342 		kmem_free(ian, sizeof (itx_async_node_t));
2343 	}
2344 	avl_destroy(t);
2345 
2346 	kmem_free(itxs, sizeof (itxs_t));
2347 }
2348 
2349 static int
2350 zil_aitx_compare(const void *x1, const void *x2)
2351 {
2352 	const uint64_t o1 = ((itx_async_node_t *)x1)->ia_foid;
2353 	const uint64_t o2 = ((itx_async_node_t *)x2)->ia_foid;
2354 
2355 	return (TREE_CMP(o1, o2));
2356 }
2357 
2358 /*
2359  * Remove all async itx with the given oid.
2360  */
2361 void
2362 zil_remove_async(zilog_t *zilog, uint64_t oid)
2363 {
2364 	uint64_t otxg, txg;
2365 	itx_async_node_t *ian;
2366 	avl_tree_t *t;
2367 	avl_index_t where;
2368 	list_t clean_list;
2369 	itx_t *itx;
2370 
2371 	ASSERT(oid != 0);
2372 	list_create(&clean_list, sizeof (itx_t), offsetof(itx_t, itx_node));
2373 
2374 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2375 		otxg = ZILTEST_TXG;
2376 	else
2377 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2378 
2379 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2380 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2381 
2382 		mutex_enter(&itxg->itxg_lock);
2383 		if (itxg->itxg_txg != txg) {
2384 			mutex_exit(&itxg->itxg_lock);
2385 			continue;
2386 		}
2387 
2388 		/*
2389 		 * Locate the object node and append its list.
2390 		 */
2391 		t = &itxg->itxg_itxs->i_async_tree;
2392 		ian = avl_find(t, &oid, &where);
2393 		if (ian != NULL)
2394 			list_move_tail(&clean_list, &ian->ia_list);
2395 		mutex_exit(&itxg->itxg_lock);
2396 	}
2397 	while ((itx = list_remove_head(&clean_list)) != NULL) {
2398 		/* commit itxs should never be on the async lists. */
2399 		ASSERT3U(itx->itx_lr.lrc_txtype, !=, TX_COMMIT);
2400 		zil_itx_destroy(itx);
2401 	}
2402 	list_destroy(&clean_list);
2403 }
2404 
2405 void
2406 zil_itx_assign(zilog_t *zilog, itx_t *itx, dmu_tx_t *tx)
2407 {
2408 	uint64_t txg;
2409 	itxg_t *itxg;
2410 	itxs_t *itxs, *clean = NULL;
2411 
2412 	/*
2413 	 * Ensure the data of a renamed file is committed before the rename.
2414 	 */
2415 	if ((itx->itx_lr.lrc_txtype & ~TX_CI) == TX_RENAME)
2416 		zil_async_to_sync(zilog, itx->itx_oid);
2417 
2418 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX)
2419 		txg = ZILTEST_TXG;
2420 	else
2421 		txg = dmu_tx_get_txg(tx);
2422 
2423 	itxg = &zilog->zl_itxg[txg & TXG_MASK];
2424 	mutex_enter(&itxg->itxg_lock);
2425 	itxs = itxg->itxg_itxs;
2426 	if (itxg->itxg_txg != txg) {
2427 		if (itxs != NULL) {
2428 			/*
2429 			 * The zil_clean callback hasn't got around to cleaning
2430 			 * this itxg. Save the itxs for release below.
2431 			 * This should be rare.
2432 			 */
2433 			zfs_dbgmsg("zil_itx_assign: missed itx cleanup for "
2434 			    "txg %llu", (u_longlong_t)itxg->itxg_txg);
2435 			clean = itxg->itxg_itxs;
2436 		}
2437 		itxg->itxg_txg = txg;
2438 		itxs = itxg->itxg_itxs = kmem_zalloc(sizeof (itxs_t),
2439 		    KM_SLEEP);
2440 
2441 		list_create(&itxs->i_sync_list, sizeof (itx_t),
2442 		    offsetof(itx_t, itx_node));
2443 		avl_create(&itxs->i_async_tree, zil_aitx_compare,
2444 		    sizeof (itx_async_node_t),
2445 		    offsetof(itx_async_node_t, ia_node));
2446 	}
2447 	if (itx->itx_sync) {
2448 		list_insert_tail(&itxs->i_sync_list, itx);
2449 	} else {
2450 		avl_tree_t *t = &itxs->i_async_tree;
2451 		uint64_t foid =
2452 		    LR_FOID_GET_OBJ(((lr_ooo_t *)&itx->itx_lr)->lr_foid);
2453 		itx_async_node_t *ian;
2454 		avl_index_t where;
2455 
2456 		ian = avl_find(t, &foid, &where);
2457 		if (ian == NULL) {
2458 			ian = kmem_alloc(sizeof (itx_async_node_t),
2459 			    KM_SLEEP);
2460 			list_create(&ian->ia_list, sizeof (itx_t),
2461 			    offsetof(itx_t, itx_node));
2462 			ian->ia_foid = foid;
2463 			avl_insert(t, ian, where);
2464 		}
2465 		list_insert_tail(&ian->ia_list, itx);
2466 	}
2467 
2468 	itx->itx_lr.lrc_txg = dmu_tx_get_txg(tx);
2469 
2470 	/*
2471 	 * We don't want to dirty the ZIL using ZILTEST_TXG, because
2472 	 * zil_clean() will never be called using ZILTEST_TXG. Thus, we
2473 	 * need to be careful to always dirty the ZIL using the "real"
2474 	 * TXG (not itxg_txg) even when the SPA is frozen.
2475 	 */
2476 	zilog_dirty(zilog, dmu_tx_get_txg(tx));
2477 	mutex_exit(&itxg->itxg_lock);
2478 
2479 	/* Release the old itxs now we've dropped the lock */
2480 	if (clean != NULL)
2481 		zil_itxg_clean(clean);
2482 }
2483 
2484 /*
2485  * If there are any in-memory intent log transactions which have now been
2486  * synced then start up a taskq to free them. We should only do this after we
2487  * have written out the uberblocks (i.e. txg has been committed) so that
2488  * don't inadvertently clean out in-memory log records that would be required
2489  * by zil_commit().
2490  */
2491 void
2492 zil_clean(zilog_t *zilog, uint64_t synced_txg)
2493 {
2494 	itxg_t *itxg = &zilog->zl_itxg[synced_txg & TXG_MASK];
2495 	itxs_t *clean_me;
2496 
2497 	ASSERT3U(synced_txg, <, ZILTEST_TXG);
2498 
2499 	mutex_enter(&itxg->itxg_lock);
2500 	if (itxg->itxg_itxs == NULL || itxg->itxg_txg == ZILTEST_TXG) {
2501 		mutex_exit(&itxg->itxg_lock);
2502 		return;
2503 	}
2504 	ASSERT3U(itxg->itxg_txg, <=, synced_txg);
2505 	ASSERT3U(itxg->itxg_txg, !=, 0);
2506 	clean_me = itxg->itxg_itxs;
2507 	itxg->itxg_itxs = NULL;
2508 	itxg->itxg_txg = 0;
2509 	mutex_exit(&itxg->itxg_lock);
2510 	/*
2511 	 * Preferably start a task queue to free up the old itxs but
2512 	 * if taskq_dispatch can't allocate resources to do that then
2513 	 * free it in-line. This should be rare. Note, using TQ_SLEEP
2514 	 * created a bad performance problem.
2515 	 */
2516 	ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
2517 	ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
2518 	taskqid_t id = taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
2519 	    zil_itxg_clean, clean_me, TQ_NOSLEEP);
2520 	if (id == TASKQID_INVALID)
2521 		zil_itxg_clean(clean_me);
2522 }
2523 
2524 /*
2525  * This function will traverse the queue of itxs that need to be
2526  * committed, and move them onto the ZIL's zl_itx_commit_list.
2527  */
2528 static uint64_t
2529 zil_get_commit_list(zilog_t *zilog)
2530 {
2531 	uint64_t otxg, txg, wtxg = 0;
2532 	list_t *commit_list = &zilog->zl_itx_commit_list;
2533 
2534 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2535 
2536 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2537 		otxg = ZILTEST_TXG;
2538 	else
2539 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2540 
2541 	/*
2542 	 * This is inherently racy, since there is nothing to prevent
2543 	 * the last synced txg from changing. That's okay since we'll
2544 	 * only commit things in the future.
2545 	 */
2546 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2547 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2548 
2549 		mutex_enter(&itxg->itxg_lock);
2550 		if (itxg->itxg_txg != txg) {
2551 			mutex_exit(&itxg->itxg_lock);
2552 			continue;
2553 		}
2554 
2555 		/*
2556 		 * If we're adding itx records to the zl_itx_commit_list,
2557 		 * then the zil better be dirty in this "txg". We can assert
2558 		 * that here since we're holding the itxg_lock which will
2559 		 * prevent spa_sync from cleaning it. Once we add the itxs
2560 		 * to the zl_itx_commit_list we must commit it to disk even
2561 		 * if it's unnecessary (i.e. the txg was synced).
2562 		 */
2563 		ASSERT(zilog_is_dirty_in_txg(zilog, txg) ||
2564 		    spa_freeze_txg(zilog->zl_spa) != UINT64_MAX);
2565 		list_t *sync_list = &itxg->itxg_itxs->i_sync_list;
2566 		if (unlikely(zilog->zl_suspend > 0)) {
2567 			/*
2568 			 * ZIL was just suspended, but we lost the race.
2569 			 * Allow all earlier itxs to be committed, but ask
2570 			 * caller to do txg_wait_synced(txg) for any new.
2571 			 */
2572 			if (!list_is_empty(sync_list))
2573 				wtxg = MAX(wtxg, txg);
2574 		} else {
2575 			list_move_tail(commit_list, sync_list);
2576 		}
2577 
2578 		mutex_exit(&itxg->itxg_lock);
2579 	}
2580 	return (wtxg);
2581 }
2582 
2583 /*
2584  * Move the async itxs for a specified object to commit into sync lists.
2585  */
2586 void
2587 zil_async_to_sync(zilog_t *zilog, uint64_t foid)
2588 {
2589 	uint64_t otxg, txg;
2590 	itx_async_node_t *ian;
2591 	avl_tree_t *t;
2592 	avl_index_t where;
2593 
2594 	if (spa_freeze_txg(zilog->zl_spa) != UINT64_MAX) /* ziltest support */
2595 		otxg = ZILTEST_TXG;
2596 	else
2597 		otxg = spa_last_synced_txg(zilog->zl_spa) + 1;
2598 
2599 	/*
2600 	 * This is inherently racy, since there is nothing to prevent
2601 	 * the last synced txg from changing.
2602 	 */
2603 	for (txg = otxg; txg < (otxg + TXG_CONCURRENT_STATES); txg++) {
2604 		itxg_t *itxg = &zilog->zl_itxg[txg & TXG_MASK];
2605 
2606 		mutex_enter(&itxg->itxg_lock);
2607 		if (itxg->itxg_txg != txg) {
2608 			mutex_exit(&itxg->itxg_lock);
2609 			continue;
2610 		}
2611 
2612 		/*
2613 		 * If a foid is specified then find that node and append its
2614 		 * list. Otherwise walk the tree appending all the lists
2615 		 * to the sync list. We add to the end rather than the
2616 		 * beginning to ensure the create has happened.
2617 		 */
2618 		t = &itxg->itxg_itxs->i_async_tree;
2619 		if (foid != 0) {
2620 			ian = avl_find(t, &foid, &where);
2621 			if (ian != NULL) {
2622 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2623 				    &ian->ia_list);
2624 			}
2625 		} else {
2626 			void *cookie = NULL;
2627 
2628 			while ((ian = avl_destroy_nodes(t, &cookie)) != NULL) {
2629 				list_move_tail(&itxg->itxg_itxs->i_sync_list,
2630 				    &ian->ia_list);
2631 				list_destroy(&ian->ia_list);
2632 				kmem_free(ian, sizeof (itx_async_node_t));
2633 			}
2634 		}
2635 		mutex_exit(&itxg->itxg_lock);
2636 	}
2637 }
2638 
2639 /*
2640  * This function will prune commit itxs that are at the head of the
2641  * commit list (it won't prune past the first non-commit itx), and
2642  * either: a) attach them to the last lwb that's still pending
2643  * completion, or b) skip them altogether.
2644  *
2645  * This is used as a performance optimization to prevent commit itxs
2646  * from generating new lwbs when it's unnecessary to do so.
2647  */
2648 static void
2649 zil_prune_commit_list(zilog_t *zilog)
2650 {
2651 	itx_t *itx;
2652 
2653 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2654 
2655 	while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) {
2656 		lr_t *lrc = &itx->itx_lr;
2657 		if (lrc->lrc_txtype != TX_COMMIT)
2658 			break;
2659 
2660 		mutex_enter(&zilog->zl_lock);
2661 
2662 		lwb_t *last_lwb = zilog->zl_last_lwb_opened;
2663 		if (last_lwb == NULL ||
2664 		    last_lwb->lwb_state == LWB_STATE_FLUSH_DONE) {
2665 			/*
2666 			 * All of the itxs this waiter was waiting on
2667 			 * must have already completed (or there were
2668 			 * never any itx's for it to wait on), so it's
2669 			 * safe to skip this waiter and mark it done.
2670 			 */
2671 			zil_commit_waiter_skip(itx->itx_private);
2672 		} else {
2673 			zil_commit_waiter_link_lwb(itx->itx_private, last_lwb);
2674 			itx->itx_private = NULL;
2675 		}
2676 
2677 		mutex_exit(&zilog->zl_lock);
2678 
2679 		list_remove(&zilog->zl_itx_commit_list, itx);
2680 		zil_itx_destroy(itx);
2681 	}
2682 
2683 	IMPLY(itx != NULL, itx->itx_lr.lrc_txtype != TX_COMMIT);
2684 }
2685 
2686 static void
2687 zil_commit_writer_stall(zilog_t *zilog)
2688 {
2689 	/*
2690 	 * When zio_alloc_zil() fails to allocate the next lwb block on
2691 	 * disk, we must call txg_wait_synced() to ensure all of the
2692 	 * lwbs in the zilog's zl_lwb_list are synced and then freed (in
2693 	 * zil_sync()), such that any subsequent ZIL writer (i.e. a call
2694 	 * to zil_process_commit_list()) will have to call zil_create(),
2695 	 * and start a new ZIL chain.
2696 	 *
2697 	 * Since zil_alloc_zil() failed, the lwb that was previously
2698 	 * issued does not have a pointer to the "next" lwb on disk.
2699 	 * Thus, if another ZIL writer thread was to allocate the "next"
2700 	 * on-disk lwb, that block could be leaked in the event of a
2701 	 * crash (because the previous lwb on-disk would not point to
2702 	 * it).
2703 	 *
2704 	 * We must hold the zilog's zl_issuer_lock while we do this, to
2705 	 * ensure no new threads enter zil_process_commit_list() until
2706 	 * all lwb's in the zl_lwb_list have been synced and freed
2707 	 * (which is achieved via the txg_wait_synced() call).
2708 	 */
2709 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2710 	txg_wait_synced(zilog->zl_dmu_pool, 0);
2711 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
2712 }
2713 
2714 /*
2715  * This function will traverse the commit list, creating new lwbs as
2716  * needed, and committing the itxs from the commit list to these newly
2717  * created lwbs. Additionally, as a new lwb is created, the previous
2718  * lwb will be issued to the zio layer to be written to disk.
2719  */
2720 static void
2721 zil_process_commit_list(zilog_t *zilog, zil_commit_waiter_t *zcw, list_t *ilwbs)
2722 {
2723 	spa_t *spa = zilog->zl_spa;
2724 	list_t nolwb_itxs;
2725 	list_t nolwb_waiters;
2726 	lwb_t *lwb, *plwb;
2727 	itx_t *itx;
2728 	boolean_t first = B_TRUE;
2729 
2730 	ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock));
2731 
2732 	/*
2733 	 * Return if there's nothing to commit before we dirty the fs by
2734 	 * calling zil_create().
2735 	 */
2736 	if (list_is_empty(&zilog->zl_itx_commit_list))
2737 		return;
2738 
2739 	list_create(&nolwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
2740 	list_create(&nolwb_waiters, sizeof (zil_commit_waiter_t),
2741 	    offsetof(zil_commit_waiter_t, zcw_node));
2742 
2743 	lwb = list_tail(&zilog->zl_lwb_list);
2744 	if (lwb == NULL) {
2745 		lwb = zil_create(zilog);
2746 	} else {
2747 		/*
2748 		 * Activate SPA_FEATURE_ZILSAXATTR for the cases where ZIL will
2749 		 * have already been created (zl_lwb_list not empty).
2750 		 */
2751 		zil_commit_activate_saxattr_feature(zilog);
2752 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2753 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2754 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2755 		first = (lwb->lwb_state != LWB_STATE_OPENED) &&
2756 		    ((plwb = list_prev(&zilog->zl_lwb_list, lwb)) == NULL ||
2757 		    plwb->lwb_state == LWB_STATE_FLUSH_DONE);
2758 	}
2759 
2760 	while ((itx = list_remove_head(&zilog->zl_itx_commit_list)) != NULL) {
2761 		lr_t *lrc = &itx->itx_lr;
2762 		uint64_t txg = lrc->lrc_txg;
2763 
2764 		ASSERT3U(txg, !=, 0);
2765 
2766 		if (lrc->lrc_txtype == TX_COMMIT) {
2767 			DTRACE_PROBE2(zil__process__commit__itx,
2768 			    zilog_t *, zilog, itx_t *, itx);
2769 		} else {
2770 			DTRACE_PROBE2(zil__process__normal__itx,
2771 			    zilog_t *, zilog, itx_t *, itx);
2772 		}
2773 
2774 		boolean_t synced = txg <= spa_last_synced_txg(spa);
2775 		boolean_t frozen = txg > spa_freeze_txg(spa);
2776 
2777 		/*
2778 		 * If the txg of this itx has already been synced out, then
2779 		 * we don't need to commit this itx to an lwb. This is
2780 		 * because the data of this itx will have already been
2781 		 * written to the main pool. This is inherently racy, and
2782 		 * it's still ok to commit an itx whose txg has already
2783 		 * been synced; this will result in a write that's
2784 		 * unnecessary, but will do no harm.
2785 		 *
2786 		 * With that said, we always want to commit TX_COMMIT itxs
2787 		 * to an lwb, regardless of whether or not that itx's txg
2788 		 * has been synced out. We do this to ensure any OPENED lwb
2789 		 * will always have at least one zil_commit_waiter_t linked
2790 		 * to the lwb.
2791 		 *
2792 		 * As a counter-example, if we skipped TX_COMMIT itx's
2793 		 * whose txg had already been synced, the following
2794 		 * situation could occur if we happened to be racing with
2795 		 * spa_sync:
2796 		 *
2797 		 * 1. We commit a non-TX_COMMIT itx to an lwb, where the
2798 		 *    itx's txg is 10 and the last synced txg is 9.
2799 		 * 2. spa_sync finishes syncing out txg 10.
2800 		 * 3. We move to the next itx in the list, it's a TX_COMMIT
2801 		 *    whose txg is 10, so we skip it rather than committing
2802 		 *    it to the lwb used in (1).
2803 		 *
2804 		 * If the itx that is skipped in (3) is the last TX_COMMIT
2805 		 * itx in the commit list, than it's possible for the lwb
2806 		 * used in (1) to remain in the OPENED state indefinitely.
2807 		 *
2808 		 * To prevent the above scenario from occurring, ensuring
2809 		 * that once an lwb is OPENED it will transition to ISSUED
2810 		 * and eventually DONE, we always commit TX_COMMIT itx's to
2811 		 * an lwb here, even if that itx's txg has already been
2812 		 * synced.
2813 		 *
2814 		 * Finally, if the pool is frozen, we _always_ commit the
2815 		 * itx.  The point of freezing the pool is to prevent data
2816 		 * from being written to the main pool via spa_sync, and
2817 		 * instead rely solely on the ZIL to persistently store the
2818 		 * data; i.e.  when the pool is frozen, the last synced txg
2819 		 * value can't be trusted.
2820 		 */
2821 		if (frozen || !synced || lrc->lrc_txtype == TX_COMMIT) {
2822 			if (lwb != NULL) {
2823 				lwb = zil_lwb_assign(zilog, lwb, itx, ilwbs);
2824 				if (lwb == NULL) {
2825 					list_insert_tail(&nolwb_itxs, itx);
2826 				} else if ((zcw->zcw_lwb != NULL &&
2827 				    zcw->zcw_lwb != lwb) || zcw->zcw_done) {
2828 					/*
2829 					 * Our lwb is done, leave the rest of
2830 					 * itx list to somebody else who care.
2831 					 */
2832 					first = B_FALSE;
2833 					break;
2834 				}
2835 			} else {
2836 				if (lrc->lrc_txtype == TX_COMMIT) {
2837 					zil_commit_waiter_link_nolwb(
2838 					    itx->itx_private, &nolwb_waiters);
2839 				}
2840 				list_insert_tail(&nolwb_itxs, itx);
2841 			}
2842 		} else {
2843 			ASSERT3S(lrc->lrc_txtype, !=, TX_COMMIT);
2844 			zil_itx_destroy(itx);
2845 		}
2846 	}
2847 
2848 	if (lwb == NULL) {
2849 		/*
2850 		 * This indicates zio_alloc_zil() failed to allocate the
2851 		 * "next" lwb on-disk. When this happens, we must stall
2852 		 * the ZIL write pipeline; see the comment within
2853 		 * zil_commit_writer_stall() for more details.
2854 		 */
2855 		while ((lwb = list_remove_head(ilwbs)) != NULL)
2856 			zil_lwb_write_issue(zilog, lwb);
2857 		zil_commit_writer_stall(zilog);
2858 
2859 		/*
2860 		 * Additionally, we have to signal and mark the "nolwb"
2861 		 * waiters as "done" here, since without an lwb, we
2862 		 * can't do this via zil_lwb_flush_vdevs_done() like
2863 		 * normal.
2864 		 */
2865 		zil_commit_waiter_t *zcw;
2866 		while ((zcw = list_remove_head(&nolwb_waiters)) != NULL)
2867 			zil_commit_waiter_skip(zcw);
2868 
2869 		/*
2870 		 * And finally, we have to destroy the itx's that
2871 		 * couldn't be committed to an lwb; this will also call
2872 		 * the itx's callback if one exists for the itx.
2873 		 */
2874 		while ((itx = list_remove_head(&nolwb_itxs)) != NULL)
2875 			zil_itx_destroy(itx);
2876 	} else {
2877 		ASSERT(list_is_empty(&nolwb_waiters));
2878 		ASSERT3P(lwb, !=, NULL);
2879 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
2880 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE);
2881 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE);
2882 
2883 		/*
2884 		 * At this point, the ZIL block pointed at by the "lwb"
2885 		 * variable is in one of the following states: "closed"
2886 		 * or "open".
2887 		 *
2888 		 * If it's "closed", then no itxs have been committed to
2889 		 * it, so there's no point in issuing its zio (i.e. it's
2890 		 * "empty").
2891 		 *
2892 		 * If it's "open", then it contains one or more itxs that
2893 		 * eventually need to be committed to stable storage. In
2894 		 * this case we intentionally do not issue the lwb's zio
2895 		 * to disk yet, and instead rely on one of the following
2896 		 * two mechanisms for issuing the zio:
2897 		 *
2898 		 * 1. Ideally, there will be more ZIL activity occurring
2899 		 * on the system, such that this function will be
2900 		 * immediately called again (not necessarily by the same
2901 		 * thread) and this lwb's zio will be issued via
2902 		 * zil_lwb_assign(). This way, the lwb is guaranteed to
2903 		 * be "full" when it is issued to disk, and we'll make
2904 		 * use of the lwb's size the best we can.
2905 		 *
2906 		 * 2. If there isn't sufficient ZIL activity occurring on
2907 		 * the system, such that this lwb's zio isn't issued via
2908 		 * zil_lwb_assign(), zil_commit_waiter() will issue the
2909 		 * lwb's zio. If this occurs, the lwb is not guaranteed
2910 		 * to be "full" by the time its zio is issued, and means
2911 		 * the size of the lwb was "too large" given the amount
2912 		 * of ZIL activity occurring on the system at that time.
2913 		 *
2914 		 * We do this for a couple of reasons:
2915 		 *
2916 		 * 1. To try and reduce the number of IOPs needed to
2917 		 * write the same number of itxs. If an lwb has space
2918 		 * available in its buffer for more itxs, and more itxs
2919 		 * will be committed relatively soon (relative to the
2920 		 * latency of performing a write), then it's beneficial
2921 		 * to wait for these "next" itxs. This way, more itxs
2922 		 * can be committed to stable storage with fewer writes.
2923 		 *
2924 		 * 2. To try and use the largest lwb block size that the
2925 		 * incoming rate of itxs can support. Again, this is to
2926 		 * try and pack as many itxs into as few lwbs as
2927 		 * possible, without significantly impacting the latency
2928 		 * of each individual itx.
2929 		 *
2930 		 * If we had no already running or open LWBs, it can be
2931 		 * the workload is single-threaded.  And if the ZIL write
2932 		 * latency is very small or if the LWB is almost full, it
2933 		 * may be cheaper to bypass the delay.
2934 		 */
2935 		if (lwb->lwb_state == LWB_STATE_OPENED && first) {
2936 			hrtime_t sleep = zilog->zl_last_lwb_latency *
2937 			    zfs_commit_timeout_pct / 100;
2938 			if (sleep < zil_min_commit_timeout ||
2939 			    lwb->lwb_sz - lwb->lwb_nused < lwb->lwb_sz / 8) {
2940 				list_insert_tail(ilwbs, lwb);
2941 				lwb = zil_lwb_write_close(zilog, lwb);
2942 				zilog->zl_cur_used = 0;
2943 				if (lwb == NULL) {
2944 					while ((lwb = list_remove_head(ilwbs))
2945 					    != NULL)
2946 						zil_lwb_write_issue(zilog, lwb);
2947 					zil_commit_writer_stall(zilog);
2948 				}
2949 			}
2950 		}
2951 	}
2952 }
2953 
2954 /*
2955  * This function is responsible for ensuring the passed in commit waiter
2956  * (and associated commit itx) is committed to an lwb. If the waiter is
2957  * not already committed to an lwb, all itxs in the zilog's queue of
2958  * itxs will be processed. The assumption is the passed in waiter's
2959  * commit itx will found in the queue just like the other non-commit
2960  * itxs, such that when the entire queue is processed, the waiter will
2961  * have been committed to an lwb.
2962  *
2963  * The lwb associated with the passed in waiter is not guaranteed to
2964  * have been issued by the time this function completes. If the lwb is
2965  * not issued, we rely on future calls to zil_commit_writer() to issue
2966  * the lwb, or the timeout mechanism found in zil_commit_waiter().
2967  */
2968 static uint64_t
2969 zil_commit_writer(zilog_t *zilog, zil_commit_waiter_t *zcw)
2970 {
2971 	list_t ilwbs;
2972 	lwb_t *lwb;
2973 	uint64_t wtxg = 0;
2974 
2975 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
2976 	ASSERT(spa_writeable(zilog->zl_spa));
2977 
2978 	list_create(&ilwbs, sizeof (lwb_t), offsetof(lwb_t, lwb_issue_node));
2979 	mutex_enter(&zilog->zl_issuer_lock);
2980 
2981 	if (zcw->zcw_lwb != NULL || zcw->zcw_done) {
2982 		/*
2983 		 * It's possible that, while we were waiting to acquire
2984 		 * the "zl_issuer_lock", another thread committed this
2985 		 * waiter to an lwb. If that occurs, we bail out early,
2986 		 * without processing any of the zilog's queue of itxs.
2987 		 *
2988 		 * On certain workloads and system configurations, the
2989 		 * "zl_issuer_lock" can become highly contended. In an
2990 		 * attempt to reduce this contention, we immediately drop
2991 		 * the lock if the waiter has already been processed.
2992 		 *
2993 		 * We've measured this optimization to reduce CPU spent
2994 		 * contending on this lock by up to 5%, using a system
2995 		 * with 32 CPUs, low latency storage (~50 usec writes),
2996 		 * and 1024 threads performing sync writes.
2997 		 */
2998 		goto out;
2999 	}
3000 
3001 	ZIL_STAT_BUMP(zilog, zil_commit_writer_count);
3002 
3003 	wtxg = zil_get_commit_list(zilog);
3004 	zil_prune_commit_list(zilog);
3005 	zil_process_commit_list(zilog, zcw, &ilwbs);
3006 
3007 out:
3008 	mutex_exit(&zilog->zl_issuer_lock);
3009 	while ((lwb = list_remove_head(&ilwbs)) != NULL)
3010 		zil_lwb_write_issue(zilog, lwb);
3011 	list_destroy(&ilwbs);
3012 	return (wtxg);
3013 }
3014 
3015 static void
3016 zil_commit_waiter_timeout(zilog_t *zilog, zil_commit_waiter_t *zcw)
3017 {
3018 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
3019 	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
3020 	ASSERT3B(zcw->zcw_done, ==, B_FALSE);
3021 
3022 	lwb_t *lwb = zcw->zcw_lwb;
3023 	ASSERT3P(lwb, !=, NULL);
3024 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_CLOSED);
3025 
3026 	/*
3027 	 * If the lwb has already been issued by another thread, we can
3028 	 * immediately return since there's no work to be done (the
3029 	 * point of this function is to issue the lwb). Additionally, we
3030 	 * do this prior to acquiring the zl_issuer_lock, to avoid
3031 	 * acquiring it when it's not necessary to do so.
3032 	 */
3033 	if (lwb->lwb_state == LWB_STATE_ISSUED ||
3034 	    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
3035 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE)
3036 		return;
3037 
3038 	/*
3039 	 * In order to call zil_lwb_write_close() we must hold the
3040 	 * zilog's "zl_issuer_lock". We can't simply acquire that lock,
3041 	 * since we're already holding the commit waiter's "zcw_lock",
3042 	 * and those two locks are acquired in the opposite order
3043 	 * elsewhere.
3044 	 */
3045 	mutex_exit(&zcw->zcw_lock);
3046 	mutex_enter(&zilog->zl_issuer_lock);
3047 	mutex_enter(&zcw->zcw_lock);
3048 
3049 	/*
3050 	 * Since we just dropped and re-acquired the commit waiter's
3051 	 * lock, we have to re-check to see if the waiter was marked
3052 	 * "done" during that process. If the waiter was marked "done",
3053 	 * the "lwb" pointer is no longer valid (it can be free'd after
3054 	 * the waiter is marked "done"), so without this check we could
3055 	 * wind up with a use-after-free error below.
3056 	 */
3057 	if (zcw->zcw_done) {
3058 		lwb = NULL;
3059 		goto out;
3060 	}
3061 
3062 	ASSERT3P(lwb, ==, zcw->zcw_lwb);
3063 
3064 	/*
3065 	 * We've already checked this above, but since we hadn't acquired
3066 	 * the zilog's zl_issuer_lock, we have to perform this check a
3067 	 * second time while holding the lock.
3068 	 *
3069 	 * We don't need to hold the zl_lock since the lwb cannot transition
3070 	 * from OPENED to ISSUED while we hold the zl_issuer_lock. The lwb
3071 	 * _can_ transition from ISSUED to DONE, but it's OK to race with
3072 	 * that transition since we treat the lwb the same, whether it's in
3073 	 * the ISSUED or DONE states.
3074 	 *
3075 	 * The important thing, is we treat the lwb differently depending on
3076 	 * if it's ISSUED or OPENED, and block any other threads that might
3077 	 * attempt to issue this lwb. For that reason we hold the
3078 	 * zl_issuer_lock when checking the lwb_state; we must not call
3079 	 * zil_lwb_write_close() if the lwb had already been issued.
3080 	 *
3081 	 * See the comment above the lwb_state_t structure definition for
3082 	 * more details on the lwb states, and locking requirements.
3083 	 */
3084 	if (lwb->lwb_state == LWB_STATE_ISSUED ||
3085 	    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
3086 	    lwb->lwb_state == LWB_STATE_FLUSH_DONE) {
3087 		lwb = NULL;
3088 		goto out;
3089 	}
3090 
3091 	ASSERT3S(lwb->lwb_state, ==, LWB_STATE_OPENED);
3092 
3093 	/*
3094 	 * As described in the comments above zil_commit_waiter() and
3095 	 * zil_process_commit_list(), we need to issue this lwb's zio
3096 	 * since we've reached the commit waiter's timeout and it still
3097 	 * hasn't been issued.
3098 	 */
3099 	lwb_t *nlwb = zil_lwb_write_close(zilog, lwb);
3100 
3101 	ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
3102 
3103 	/*
3104 	 * Since the lwb's zio hadn't been issued by the time this thread
3105 	 * reached its timeout, we reset the zilog's "zl_cur_used" field
3106 	 * to influence the zil block size selection algorithm.
3107 	 *
3108 	 * By having to issue the lwb's zio here, it means the size of the
3109 	 * lwb was too large, given the incoming throughput of itxs.  By
3110 	 * setting "zl_cur_used" to zero, we communicate this fact to the
3111 	 * block size selection algorithm, so it can take this information
3112 	 * into account, and potentially select a smaller size for the
3113 	 * next lwb block that is allocated.
3114 	 */
3115 	zilog->zl_cur_used = 0;
3116 
3117 	if (nlwb == NULL) {
3118 		/*
3119 		 * When zil_lwb_write_close() returns NULL, this
3120 		 * indicates zio_alloc_zil() failed to allocate the
3121 		 * "next" lwb on-disk. When this occurs, the ZIL write
3122 		 * pipeline must be stalled; see the comment within the
3123 		 * zil_commit_writer_stall() function for more details.
3124 		 *
3125 		 * We must drop the commit waiter's lock prior to
3126 		 * calling zil_commit_writer_stall() or else we can wind
3127 		 * up with the following deadlock:
3128 		 *
3129 		 * - This thread is waiting for the txg to sync while
3130 		 *   holding the waiter's lock; txg_wait_synced() is
3131 		 *   used within txg_commit_writer_stall().
3132 		 *
3133 		 * - The txg can't sync because it is waiting for this
3134 		 *   lwb's zio callback to call dmu_tx_commit().
3135 		 *
3136 		 * - The lwb's zio callback can't call dmu_tx_commit()
3137 		 *   because it's blocked trying to acquire the waiter's
3138 		 *   lock, which occurs prior to calling dmu_tx_commit()
3139 		 */
3140 		mutex_exit(&zcw->zcw_lock);
3141 		zil_lwb_write_issue(zilog, lwb);
3142 		lwb = NULL;
3143 		zil_commit_writer_stall(zilog);
3144 		mutex_enter(&zcw->zcw_lock);
3145 	}
3146 
3147 out:
3148 	mutex_exit(&zilog->zl_issuer_lock);
3149 	if (lwb)
3150 		zil_lwb_write_issue(zilog, lwb);
3151 	ASSERT(MUTEX_HELD(&zcw->zcw_lock));
3152 }
3153 
3154 /*
3155  * This function is responsible for performing the following two tasks:
3156  *
3157  * 1. its primary responsibility is to block until the given "commit
3158  *    waiter" is considered "done".
3159  *
3160  * 2. its secondary responsibility is to issue the zio for the lwb that
3161  *    the given "commit waiter" is waiting on, if this function has
3162  *    waited "long enough" and the lwb is still in the "open" state.
3163  *
3164  * Given a sufficient amount of itxs being generated and written using
3165  * the ZIL, the lwb's zio will be issued via the zil_lwb_assign()
3166  * function. If this does not occur, this secondary responsibility will
3167  * ensure the lwb is issued even if there is not other synchronous
3168  * activity on the system.
3169  *
3170  * For more details, see zil_process_commit_list(); more specifically,
3171  * the comment at the bottom of that function.
3172  */
3173 static void
3174 zil_commit_waiter(zilog_t *zilog, zil_commit_waiter_t *zcw)
3175 {
3176 	ASSERT(!MUTEX_HELD(&zilog->zl_lock));
3177 	ASSERT(!MUTEX_HELD(&zilog->zl_issuer_lock));
3178 	ASSERT(spa_writeable(zilog->zl_spa));
3179 
3180 	mutex_enter(&zcw->zcw_lock);
3181 
3182 	/*
3183 	 * The timeout is scaled based on the lwb latency to avoid
3184 	 * significantly impacting the latency of each individual itx.
3185 	 * For more details, see the comment at the bottom of the
3186 	 * zil_process_commit_list() function.
3187 	 */
3188 	int pct = MAX(zfs_commit_timeout_pct, 1);
3189 	hrtime_t sleep = (zilog->zl_last_lwb_latency * pct) / 100;
3190 	hrtime_t wakeup = gethrtime() + sleep;
3191 	boolean_t timedout = B_FALSE;
3192 
3193 	while (!zcw->zcw_done) {
3194 		ASSERT(MUTEX_HELD(&zcw->zcw_lock));
3195 
3196 		lwb_t *lwb = zcw->zcw_lwb;
3197 
3198 		/*
3199 		 * Usually, the waiter will have a non-NULL lwb field here,
3200 		 * but it's possible for it to be NULL as a result of
3201 		 * zil_commit() racing with spa_sync().
3202 		 *
3203 		 * When zil_clean() is called, it's possible for the itxg
3204 		 * list (which may be cleaned via a taskq) to contain
3205 		 * commit itxs. When this occurs, the commit waiters linked
3206 		 * off of these commit itxs will not be committed to an
3207 		 * lwb.  Additionally, these commit waiters will not be
3208 		 * marked done until zil_commit_waiter_skip() is called via
3209 		 * zil_itxg_clean().
3210 		 *
3211 		 * Thus, it's possible for this commit waiter (i.e. the
3212 		 * "zcw" variable) to be found in this "in between" state;
3213 		 * where it's "zcw_lwb" field is NULL, and it hasn't yet
3214 		 * been skipped, so it's "zcw_done" field is still B_FALSE.
3215 		 */
3216 		IMPLY(lwb != NULL, lwb->lwb_state != LWB_STATE_CLOSED);
3217 
3218 		if (lwb != NULL && lwb->lwb_state == LWB_STATE_OPENED) {
3219 			ASSERT3B(timedout, ==, B_FALSE);
3220 
3221 			/*
3222 			 * If the lwb hasn't been issued yet, then we
3223 			 * need to wait with a timeout, in case this
3224 			 * function needs to issue the lwb after the
3225 			 * timeout is reached; responsibility (2) from
3226 			 * the comment above this function.
3227 			 */
3228 			int rc = cv_timedwait_hires(&zcw->zcw_cv,
3229 			    &zcw->zcw_lock, wakeup, USEC2NSEC(1),
3230 			    CALLOUT_FLAG_ABSOLUTE);
3231 
3232 			if (rc != -1 || zcw->zcw_done)
3233 				continue;
3234 
3235 			timedout = B_TRUE;
3236 			zil_commit_waiter_timeout(zilog, zcw);
3237 
3238 			if (!zcw->zcw_done) {
3239 				/*
3240 				 * If the commit waiter has already been
3241 				 * marked "done", it's possible for the
3242 				 * waiter's lwb structure to have already
3243 				 * been freed.  Thus, we can only reliably
3244 				 * make these assertions if the waiter
3245 				 * isn't done.
3246 				 */
3247 				ASSERT3P(lwb, ==, zcw->zcw_lwb);
3248 				ASSERT3S(lwb->lwb_state, !=, LWB_STATE_OPENED);
3249 			}
3250 		} else {
3251 			/*
3252 			 * If the lwb isn't open, then it must have already
3253 			 * been issued. In that case, there's no need to
3254 			 * use a timeout when waiting for the lwb to
3255 			 * complete.
3256 			 *
3257 			 * Additionally, if the lwb is NULL, the waiter
3258 			 * will soon be signaled and marked done via
3259 			 * zil_clean() and zil_itxg_clean(), so no timeout
3260 			 * is required.
3261 			 */
3262 
3263 			IMPLY(lwb != NULL,
3264 			    lwb->lwb_state == LWB_STATE_ISSUED ||
3265 			    lwb->lwb_state == LWB_STATE_WRITE_DONE ||
3266 			    lwb->lwb_state == LWB_STATE_FLUSH_DONE);
3267 			cv_wait(&zcw->zcw_cv, &zcw->zcw_lock);
3268 		}
3269 	}
3270 
3271 	mutex_exit(&zcw->zcw_lock);
3272 }
3273 
3274 static zil_commit_waiter_t *
3275 zil_alloc_commit_waiter(void)
3276 {
3277 	zil_commit_waiter_t *zcw = kmem_cache_alloc(zil_zcw_cache, KM_SLEEP);
3278 
3279 	cv_init(&zcw->zcw_cv, NULL, CV_DEFAULT, NULL);
3280 	mutex_init(&zcw->zcw_lock, NULL, MUTEX_DEFAULT, NULL);
3281 	list_link_init(&zcw->zcw_node);
3282 	zcw->zcw_lwb = NULL;
3283 	zcw->zcw_done = B_FALSE;
3284 	zcw->zcw_zio_error = 0;
3285 
3286 	return (zcw);
3287 }
3288 
3289 static void
3290 zil_free_commit_waiter(zil_commit_waiter_t *zcw)
3291 {
3292 	ASSERT(!list_link_active(&zcw->zcw_node));
3293 	ASSERT3P(zcw->zcw_lwb, ==, NULL);
3294 	ASSERT3B(zcw->zcw_done, ==, B_TRUE);
3295 	mutex_destroy(&zcw->zcw_lock);
3296 	cv_destroy(&zcw->zcw_cv);
3297 	kmem_cache_free(zil_zcw_cache, zcw);
3298 }
3299 
3300 /*
3301  * This function is used to create a TX_COMMIT itx and assign it. This
3302  * way, it will be linked into the ZIL's list of synchronous itxs, and
3303  * then later committed to an lwb (or skipped) when
3304  * zil_process_commit_list() is called.
3305  */
3306 static void
3307 zil_commit_itx_assign(zilog_t *zilog, zil_commit_waiter_t *zcw)
3308 {
3309 	dmu_tx_t *tx = dmu_tx_create(zilog->zl_os);
3310 
3311 	/*
3312 	 * Since we are not going to create any new dirty data, and we
3313 	 * can even help with clearing the existing dirty data, we
3314 	 * should not be subject to the dirty data based delays. We
3315 	 * use TXG_NOTHROTTLE to bypass the delay mechanism.
3316 	 */
3317 	VERIFY0(dmu_tx_assign(tx, TXG_WAIT | TXG_NOTHROTTLE));
3318 
3319 	itx_t *itx = zil_itx_create(TX_COMMIT, sizeof (lr_t));
3320 	itx->itx_sync = B_TRUE;
3321 	itx->itx_private = zcw;
3322 
3323 	zil_itx_assign(zilog, itx, tx);
3324 
3325 	dmu_tx_commit(tx);
3326 }
3327 
3328 /*
3329  * Commit ZFS Intent Log transactions (itxs) to stable storage.
3330  *
3331  * When writing ZIL transactions to the on-disk representation of the
3332  * ZIL, the itxs are committed to a Log Write Block (lwb). Multiple
3333  * itxs can be committed to a single lwb. Once a lwb is written and
3334  * committed to stable storage (i.e. the lwb is written, and vdevs have
3335  * been flushed), each itx that was committed to that lwb is also
3336  * considered to be committed to stable storage.
3337  *
3338  * When an itx is committed to an lwb, the log record (lr_t) contained
3339  * by the itx is copied into the lwb's zio buffer, and once this buffer
3340  * is written to disk, it becomes an on-disk ZIL block.
3341  *
3342  * As itxs are generated, they're inserted into the ZIL's queue of
3343  * uncommitted itxs. The semantics of zil_commit() are such that it will
3344  * block until all itxs that were in the queue when it was called, are
3345  * committed to stable storage.
3346  *
3347  * If "foid" is zero, this means all "synchronous" and "asynchronous"
3348  * itxs, for all objects in the dataset, will be committed to stable
3349  * storage prior to zil_commit() returning. If "foid" is non-zero, all
3350  * "synchronous" itxs for all objects, but only "asynchronous" itxs
3351  * that correspond to the foid passed in, will be committed to stable
3352  * storage prior to zil_commit() returning.
3353  *
3354  * Generally speaking, when zil_commit() is called, the consumer doesn't
3355  * actually care about _all_ of the uncommitted itxs. Instead, they're
3356  * simply trying to waiting for a specific itx to be committed to disk,
3357  * but the interface(s) for interacting with the ZIL don't allow such
3358  * fine-grained communication. A better interface would allow a consumer
3359  * to create and assign an itx, and then pass a reference to this itx to
3360  * zil_commit(); such that zil_commit() would return as soon as that
3361  * specific itx was committed to disk (instead of waiting for _all_
3362  * itxs to be committed).
3363  *
3364  * When a thread calls zil_commit() a special "commit itx" will be
3365  * generated, along with a corresponding "waiter" for this commit itx.
3366  * zil_commit() will wait on this waiter's CV, such that when the waiter
3367  * is marked done, and signaled, zil_commit() will return.
3368  *
3369  * This commit itx is inserted into the queue of uncommitted itxs. This
3370  * provides an easy mechanism for determining which itxs were in the
3371  * queue prior to zil_commit() having been called, and which itxs were
3372  * added after zil_commit() was called.
3373  *
3374  * The commit itx is special; it doesn't have any on-disk representation.
3375  * When a commit itx is "committed" to an lwb, the waiter associated
3376  * with it is linked onto the lwb's list of waiters. Then, when that lwb
3377  * completes, each waiter on the lwb's list is marked done and signaled
3378  * -- allowing the thread waiting on the waiter to return from zil_commit().
3379  *
3380  * It's important to point out a few critical factors that allow us
3381  * to make use of the commit itxs, commit waiters, per-lwb lists of
3382  * commit waiters, and zio completion callbacks like we're doing:
3383  *
3384  *   1. The list of waiters for each lwb is traversed, and each commit
3385  *      waiter is marked "done" and signaled, in the zio completion
3386  *      callback of the lwb's zio[*].
3387  *
3388  *      * Actually, the waiters are signaled in the zio completion
3389  *        callback of the root zio for the DKIOCFLUSHWRITECACHE commands
3390  *        that are sent to the vdevs upon completion of the lwb zio.
3391  *
3392  *   2. When the itxs are inserted into the ZIL's queue of uncommitted
3393  *      itxs, the order in which they are inserted is preserved[*]; as
3394  *      itxs are added to the queue, they are added to the tail of
3395  *      in-memory linked lists.
3396  *
3397  *      When committing the itxs to lwbs (to be written to disk), they
3398  *      are committed in the same order in which the itxs were added to
3399  *      the uncommitted queue's linked list(s); i.e. the linked list of
3400  *      itxs to commit is traversed from head to tail, and each itx is
3401  *      committed to an lwb in that order.
3402  *
3403  *      * To clarify:
3404  *
3405  *        - the order of "sync" itxs is preserved w.r.t. other
3406  *          "sync" itxs, regardless of the corresponding objects.
3407  *        - the order of "async" itxs is preserved w.r.t. other
3408  *          "async" itxs corresponding to the same object.
3409  *        - the order of "async" itxs is *not* preserved w.r.t. other
3410  *          "async" itxs corresponding to different objects.
3411  *        - the order of "sync" itxs w.r.t. "async" itxs (or vice
3412  *          versa) is *not* preserved, even for itxs that correspond
3413  *          to the same object.
3414  *
3415  *      For more details, see: zil_itx_assign(), zil_async_to_sync(),
3416  *      zil_get_commit_list(), and zil_process_commit_list().
3417  *
3418  *   3. The lwbs represent a linked list of blocks on disk. Thus, any
3419  *      lwb cannot be considered committed to stable storage, until its
3420  *      "previous" lwb is also committed to stable storage. This fact,
3421  *      coupled with the fact described above, means that itxs are
3422  *      committed in (roughly) the order in which they were generated.
3423  *      This is essential because itxs are dependent on prior itxs.
3424  *      Thus, we *must not* deem an itx as being committed to stable
3425  *      storage, until *all* prior itxs have also been committed to
3426  *      stable storage.
3427  *
3428  *      To enforce this ordering of lwb zio's, while still leveraging as
3429  *      much of the underlying storage performance as possible, we rely
3430  *      on two fundamental concepts:
3431  *
3432  *          1. The creation and issuance of lwb zio's is protected by
3433  *             the zilog's "zl_issuer_lock", which ensures only a single
3434  *             thread is creating and/or issuing lwb's at a time
3435  *          2. The "previous" lwb is a child of the "current" lwb
3436  *             (leveraging the zio parent-child dependency graph)
3437  *
3438  *      By relying on this parent-child zio relationship, we can have
3439  *      many lwb zio's concurrently issued to the underlying storage,
3440  *      but the order in which they complete will be the same order in
3441  *      which they were created.
3442  */
3443 void
3444 zil_commit(zilog_t *zilog, uint64_t foid)
3445 {
3446 	/*
3447 	 * We should never attempt to call zil_commit on a snapshot for
3448 	 * a couple of reasons:
3449 	 *
3450 	 * 1. A snapshot may never be modified, thus it cannot have any
3451 	 *    in-flight itxs that would have modified the dataset.
3452 	 *
3453 	 * 2. By design, when zil_commit() is called, a commit itx will
3454 	 *    be assigned to this zilog; as a result, the zilog will be
3455 	 *    dirtied. We must not dirty the zilog of a snapshot; there's
3456 	 *    checks in the code that enforce this invariant, and will
3457 	 *    cause a panic if it's not upheld.
3458 	 */
3459 	ASSERT3B(dmu_objset_is_snapshot(zilog->zl_os), ==, B_FALSE);
3460 
3461 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
3462 		return;
3463 
3464 	if (!spa_writeable(zilog->zl_spa)) {
3465 		/*
3466 		 * If the SPA is not writable, there should never be any
3467 		 * pending itxs waiting to be committed to disk. If that
3468 		 * weren't true, we'd skip writing those itxs out, and
3469 		 * would break the semantics of zil_commit(); thus, we're
3470 		 * verifying that truth before we return to the caller.
3471 		 */
3472 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3473 		ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3474 		for (int i = 0; i < TXG_SIZE; i++)
3475 			ASSERT3P(zilog->zl_itxg[i].itxg_itxs, ==, NULL);
3476 		return;
3477 	}
3478 
3479 	/*
3480 	 * If the ZIL is suspended, we don't want to dirty it by calling
3481 	 * zil_commit_itx_assign() below, nor can we write out
3482 	 * lwbs like would be done in zil_commit_write(). Thus, we
3483 	 * simply rely on txg_wait_synced() to maintain the necessary
3484 	 * semantics, and avoid calling those functions altogether.
3485 	 */
3486 	if (zilog->zl_suspend > 0) {
3487 		txg_wait_synced(zilog->zl_dmu_pool, 0);
3488 		return;
3489 	}
3490 
3491 	zil_commit_impl(zilog, foid);
3492 }
3493 
3494 void
3495 zil_commit_impl(zilog_t *zilog, uint64_t foid)
3496 {
3497 	ZIL_STAT_BUMP(zilog, zil_commit_count);
3498 
3499 	/*
3500 	 * Move the "async" itxs for the specified foid to the "sync"
3501 	 * queues, such that they will be later committed (or skipped)
3502 	 * to an lwb when zil_process_commit_list() is called.
3503 	 *
3504 	 * Since these "async" itxs must be committed prior to this
3505 	 * call to zil_commit returning, we must perform this operation
3506 	 * before we call zil_commit_itx_assign().
3507 	 */
3508 	zil_async_to_sync(zilog, foid);
3509 
3510 	/*
3511 	 * We allocate a new "waiter" structure which will initially be
3512 	 * linked to the commit itx using the itx's "itx_private" field.
3513 	 * Since the commit itx doesn't represent any on-disk state,
3514 	 * when it's committed to an lwb, rather than copying the its
3515 	 * lr_t into the lwb's buffer, the commit itx's "waiter" will be
3516 	 * added to the lwb's list of waiters. Then, when the lwb is
3517 	 * committed to stable storage, each waiter in the lwb's list of
3518 	 * waiters will be marked "done", and signalled.
3519 	 *
3520 	 * We must create the waiter and assign the commit itx prior to
3521 	 * calling zil_commit_writer(), or else our specific commit itx
3522 	 * is not guaranteed to be committed to an lwb prior to calling
3523 	 * zil_commit_waiter().
3524 	 */
3525 	zil_commit_waiter_t *zcw = zil_alloc_commit_waiter();
3526 	zil_commit_itx_assign(zilog, zcw);
3527 
3528 	uint64_t wtxg = zil_commit_writer(zilog, zcw);
3529 	zil_commit_waiter(zilog, zcw);
3530 
3531 	if (zcw->zcw_zio_error != 0) {
3532 		/*
3533 		 * If there was an error writing out the ZIL blocks that
3534 		 * this thread is waiting on, then we fallback to
3535 		 * relying on spa_sync() to write out the data this
3536 		 * thread is waiting on. Obviously this has performance
3537 		 * implications, but the expectation is for this to be
3538 		 * an exceptional case, and shouldn't occur often.
3539 		 */
3540 		DTRACE_PROBE2(zil__commit__io__error,
3541 		    zilog_t *, zilog, zil_commit_waiter_t *, zcw);
3542 		txg_wait_synced(zilog->zl_dmu_pool, 0);
3543 	} else if (wtxg != 0) {
3544 		txg_wait_synced(zilog->zl_dmu_pool, wtxg);
3545 	}
3546 
3547 	zil_free_commit_waiter(zcw);
3548 }
3549 
3550 /*
3551  * Called in syncing context to free committed log blocks and update log header.
3552  */
3553 void
3554 zil_sync(zilog_t *zilog, dmu_tx_t *tx)
3555 {
3556 	zil_header_t *zh = zil_header_in_syncing_context(zilog);
3557 	uint64_t txg = dmu_tx_get_txg(tx);
3558 	spa_t *spa = zilog->zl_spa;
3559 	uint64_t *replayed_seq = &zilog->zl_replayed_seq[txg & TXG_MASK];
3560 	lwb_t *lwb;
3561 
3562 	/*
3563 	 * We don't zero out zl_destroy_txg, so make sure we don't try
3564 	 * to destroy it twice.
3565 	 */
3566 	if (spa_sync_pass(spa) != 1)
3567 		return;
3568 
3569 	zil_lwb_flush_wait_all(zilog, txg);
3570 
3571 	mutex_enter(&zilog->zl_lock);
3572 
3573 	ASSERT(zilog->zl_stop_sync == 0);
3574 
3575 	if (*replayed_seq != 0) {
3576 		ASSERT(zh->zh_replay_seq < *replayed_seq);
3577 		zh->zh_replay_seq = *replayed_seq;
3578 		*replayed_seq = 0;
3579 	}
3580 
3581 	if (zilog->zl_destroy_txg == txg) {
3582 		blkptr_t blk = zh->zh_log;
3583 		dsl_dataset_t *ds = dmu_objset_ds(zilog->zl_os);
3584 
3585 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3586 
3587 		memset(zh, 0, sizeof (zil_header_t));
3588 		memset(zilog->zl_replayed_seq, 0,
3589 		    sizeof (zilog->zl_replayed_seq));
3590 
3591 		if (zilog->zl_keep_first) {
3592 			/*
3593 			 * If this block was part of log chain that couldn't
3594 			 * be claimed because a device was missing during
3595 			 * zil_claim(), but that device later returns,
3596 			 * then this block could erroneously appear valid.
3597 			 * To guard against this, assign a new GUID to the new
3598 			 * log chain so it doesn't matter what blk points to.
3599 			 */
3600 			zil_init_log_chain(zilog, &blk);
3601 			zh->zh_log = blk;
3602 		} else {
3603 			/*
3604 			 * A destroyed ZIL chain can't contain any TX_SETSAXATTR
3605 			 * records. So, deactivate the feature for this dataset.
3606 			 * We activate it again when we start a new ZIL chain.
3607 			 */
3608 			if (dsl_dataset_feature_is_active(ds,
3609 			    SPA_FEATURE_ZILSAXATTR))
3610 				dsl_dataset_deactivate_feature(ds,
3611 				    SPA_FEATURE_ZILSAXATTR, tx);
3612 		}
3613 	}
3614 
3615 	while ((lwb = list_head(&zilog->zl_lwb_list)) != NULL) {
3616 		zh->zh_log = lwb->lwb_blk;
3617 		if (lwb->lwb_state != LWB_STATE_FLUSH_DONE ||
3618 		    lwb->lwb_max_txg > txg)
3619 			break;
3620 		list_remove(&zilog->zl_lwb_list, lwb);
3621 		zio_free(spa, txg, &lwb->lwb_blk);
3622 		zil_free_lwb(zilog, lwb);
3623 
3624 		/*
3625 		 * If we don't have anything left in the lwb list then
3626 		 * we've had an allocation failure and we need to zero
3627 		 * out the zil_header blkptr so that we don't end
3628 		 * up freeing the same block twice.
3629 		 */
3630 		if (list_is_empty(&zilog->zl_lwb_list))
3631 			BP_ZERO(&zh->zh_log);
3632 	}
3633 
3634 	/*
3635 	 * Remove fastwrite on any blocks that have been pre-allocated for
3636 	 * the next commit. This prevents fastwrite counter pollution by
3637 	 * unused, long-lived LWBs.
3638 	 */
3639 	for (; lwb != NULL; lwb = list_next(&zilog->zl_lwb_list, lwb)) {
3640 		if (lwb->lwb_fastwrite && !lwb->lwb_write_zio) {
3641 			metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
3642 			lwb->lwb_fastwrite = 0;
3643 		}
3644 	}
3645 
3646 	mutex_exit(&zilog->zl_lock);
3647 }
3648 
3649 static int
3650 zil_lwb_cons(void *vbuf, void *unused, int kmflag)
3651 {
3652 	(void) unused, (void) kmflag;
3653 	lwb_t *lwb = vbuf;
3654 	list_create(&lwb->lwb_itxs, sizeof (itx_t), offsetof(itx_t, itx_node));
3655 	list_create(&lwb->lwb_waiters, sizeof (zil_commit_waiter_t),
3656 	    offsetof(zil_commit_waiter_t, zcw_node));
3657 	avl_create(&lwb->lwb_vdev_tree, zil_lwb_vdev_compare,
3658 	    sizeof (zil_vdev_node_t), offsetof(zil_vdev_node_t, zv_node));
3659 	mutex_init(&lwb->lwb_vdev_lock, NULL, MUTEX_DEFAULT, NULL);
3660 	return (0);
3661 }
3662 
3663 static void
3664 zil_lwb_dest(void *vbuf, void *unused)
3665 {
3666 	(void) unused;
3667 	lwb_t *lwb = vbuf;
3668 	mutex_destroy(&lwb->lwb_vdev_lock);
3669 	avl_destroy(&lwb->lwb_vdev_tree);
3670 	list_destroy(&lwb->lwb_waiters);
3671 	list_destroy(&lwb->lwb_itxs);
3672 }
3673 
3674 void
3675 zil_init(void)
3676 {
3677 	zil_lwb_cache = kmem_cache_create("zil_lwb_cache",
3678 	    sizeof (lwb_t), 0, zil_lwb_cons, zil_lwb_dest, NULL, NULL, NULL, 0);
3679 
3680 	zil_zcw_cache = kmem_cache_create("zil_zcw_cache",
3681 	    sizeof (zil_commit_waiter_t), 0, NULL, NULL, NULL, NULL, NULL, 0);
3682 
3683 	zil_sums_init(&zil_sums_global);
3684 	zil_kstats_global = kstat_create("zfs", 0, "zil", "misc",
3685 	    KSTAT_TYPE_NAMED, sizeof (zil_stats) / sizeof (kstat_named_t),
3686 	    KSTAT_FLAG_VIRTUAL);
3687 
3688 	if (zil_kstats_global != NULL) {
3689 		zil_kstats_global->ks_data = &zil_stats;
3690 		zil_kstats_global->ks_update = zil_kstats_global_update;
3691 		zil_kstats_global->ks_private = NULL;
3692 		kstat_install(zil_kstats_global);
3693 	}
3694 }
3695 
3696 void
3697 zil_fini(void)
3698 {
3699 	kmem_cache_destroy(zil_zcw_cache);
3700 	kmem_cache_destroy(zil_lwb_cache);
3701 
3702 	if (zil_kstats_global != NULL) {
3703 		kstat_delete(zil_kstats_global);
3704 		zil_kstats_global = NULL;
3705 	}
3706 
3707 	zil_sums_fini(&zil_sums_global);
3708 }
3709 
3710 void
3711 zil_set_sync(zilog_t *zilog, uint64_t sync)
3712 {
3713 	zilog->zl_sync = sync;
3714 }
3715 
3716 void
3717 zil_set_logbias(zilog_t *zilog, uint64_t logbias)
3718 {
3719 	zilog->zl_logbias = logbias;
3720 }
3721 
3722 zilog_t *
3723 zil_alloc(objset_t *os, zil_header_t *zh_phys)
3724 {
3725 	zilog_t *zilog;
3726 
3727 	zilog = kmem_zalloc(sizeof (zilog_t), KM_SLEEP);
3728 
3729 	zilog->zl_header = zh_phys;
3730 	zilog->zl_os = os;
3731 	zilog->zl_spa = dmu_objset_spa(os);
3732 	zilog->zl_dmu_pool = dmu_objset_pool(os);
3733 	zilog->zl_destroy_txg = TXG_INITIAL - 1;
3734 	zilog->zl_logbias = dmu_objset_logbias(os);
3735 	zilog->zl_sync = dmu_objset_syncprop(os);
3736 	zilog->zl_dirty_max_txg = 0;
3737 	zilog->zl_last_lwb_opened = NULL;
3738 	zilog->zl_last_lwb_latency = 0;
3739 	zilog->zl_max_block_size = zil_maxblocksize;
3740 
3741 	mutex_init(&zilog->zl_lock, NULL, MUTEX_DEFAULT, NULL);
3742 	mutex_init(&zilog->zl_issuer_lock, NULL, MUTEX_DEFAULT, NULL);
3743 	mutex_init(&zilog->zl_lwb_io_lock, NULL, MUTEX_DEFAULT, NULL);
3744 
3745 	for (int i = 0; i < TXG_SIZE; i++) {
3746 		mutex_init(&zilog->zl_itxg[i].itxg_lock, NULL,
3747 		    MUTEX_DEFAULT, NULL);
3748 	}
3749 
3750 	list_create(&zilog->zl_lwb_list, sizeof (lwb_t),
3751 	    offsetof(lwb_t, lwb_node));
3752 
3753 	list_create(&zilog->zl_itx_commit_list, sizeof (itx_t),
3754 	    offsetof(itx_t, itx_node));
3755 
3756 	cv_init(&zilog->zl_cv_suspend, NULL, CV_DEFAULT, NULL);
3757 	cv_init(&zilog->zl_lwb_io_cv, NULL, CV_DEFAULT, NULL);
3758 
3759 	return (zilog);
3760 }
3761 
3762 void
3763 zil_free(zilog_t *zilog)
3764 {
3765 	int i;
3766 
3767 	zilog->zl_stop_sync = 1;
3768 
3769 	ASSERT0(zilog->zl_suspend);
3770 	ASSERT0(zilog->zl_suspending);
3771 
3772 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
3773 	list_destroy(&zilog->zl_lwb_list);
3774 
3775 	ASSERT(list_is_empty(&zilog->zl_itx_commit_list));
3776 	list_destroy(&zilog->zl_itx_commit_list);
3777 
3778 	for (i = 0; i < TXG_SIZE; i++) {
3779 		/*
3780 		 * It's possible for an itx to be generated that doesn't dirty
3781 		 * a txg (e.g. ztest TX_TRUNCATE). So there's no zil_clean()
3782 		 * callback to remove the entry. We remove those here.
3783 		 *
3784 		 * Also free up the ziltest itxs.
3785 		 */
3786 		if (zilog->zl_itxg[i].itxg_itxs)
3787 			zil_itxg_clean(zilog->zl_itxg[i].itxg_itxs);
3788 		mutex_destroy(&zilog->zl_itxg[i].itxg_lock);
3789 	}
3790 
3791 	mutex_destroy(&zilog->zl_issuer_lock);
3792 	mutex_destroy(&zilog->zl_lock);
3793 	mutex_destroy(&zilog->zl_lwb_io_lock);
3794 
3795 	cv_destroy(&zilog->zl_cv_suspend);
3796 	cv_destroy(&zilog->zl_lwb_io_cv);
3797 
3798 	kmem_free(zilog, sizeof (zilog_t));
3799 }
3800 
3801 /*
3802  * Open an intent log.
3803  */
3804 zilog_t *
3805 zil_open(objset_t *os, zil_get_data_t *get_data, zil_sums_t *zil_sums)
3806 {
3807 	zilog_t *zilog = dmu_objset_zil(os);
3808 
3809 	ASSERT3P(zilog->zl_get_data, ==, NULL);
3810 	ASSERT3P(zilog->zl_last_lwb_opened, ==, NULL);
3811 	ASSERT(list_is_empty(&zilog->zl_lwb_list));
3812 
3813 	zilog->zl_get_data = get_data;
3814 	zilog->zl_sums = zil_sums;
3815 
3816 	return (zilog);
3817 }
3818 
3819 /*
3820  * Close an intent log.
3821  */
3822 void
3823 zil_close(zilog_t *zilog)
3824 {
3825 	lwb_t *lwb;
3826 	uint64_t txg;
3827 
3828 	if (!dmu_objset_is_snapshot(zilog->zl_os)) {
3829 		zil_commit(zilog, 0);
3830 	} else {
3831 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3832 		ASSERT0(zilog->zl_dirty_max_txg);
3833 		ASSERT3B(zilog_is_dirty(zilog), ==, B_FALSE);
3834 	}
3835 
3836 	mutex_enter(&zilog->zl_lock);
3837 	lwb = list_tail(&zilog->zl_lwb_list);
3838 	if (lwb == NULL)
3839 		txg = zilog->zl_dirty_max_txg;
3840 	else
3841 		txg = MAX(zilog->zl_dirty_max_txg, lwb->lwb_max_txg);
3842 	mutex_exit(&zilog->zl_lock);
3843 
3844 	/*
3845 	 * zl_lwb_max_issued_txg may be larger than lwb_max_txg. It depends
3846 	 * on the time when the dmu_tx transaction is assigned in
3847 	 * zil_lwb_write_close().
3848 	 */
3849 	mutex_enter(&zilog->zl_lwb_io_lock);
3850 	txg = MAX(zilog->zl_lwb_max_issued_txg, txg);
3851 	mutex_exit(&zilog->zl_lwb_io_lock);
3852 
3853 	/*
3854 	 * We need to use txg_wait_synced() to wait until that txg is synced.
3855 	 * zil_sync() will guarantee all lwbs up to that txg have been
3856 	 * written out, flushed, and cleaned.
3857 	 */
3858 	if (txg != 0)
3859 		txg_wait_synced(zilog->zl_dmu_pool, txg);
3860 
3861 	if (zilog_is_dirty(zilog))
3862 		zfs_dbgmsg("zil (%px) is dirty, txg %llu", zilog,
3863 		    (u_longlong_t)txg);
3864 	if (txg < spa_freeze_txg(zilog->zl_spa))
3865 		VERIFY(!zilog_is_dirty(zilog));
3866 
3867 	zilog->zl_get_data = NULL;
3868 
3869 	/*
3870 	 * We should have only one lwb left on the list; remove it now.
3871 	 */
3872 	mutex_enter(&zilog->zl_lock);
3873 	lwb = list_remove_head(&zilog->zl_lwb_list);
3874 	if (lwb != NULL) {
3875 		ASSERT(list_is_empty(&zilog->zl_lwb_list));
3876 		ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED);
3877 
3878 		if (lwb->lwb_fastwrite)
3879 			metaslab_fastwrite_unmark(zilog->zl_spa, &lwb->lwb_blk);
3880 
3881 		zio_buf_free(lwb->lwb_buf, lwb->lwb_sz);
3882 		zil_free_lwb(zilog, lwb);
3883 	}
3884 	mutex_exit(&zilog->zl_lock);
3885 }
3886 
3887 static const char *suspend_tag = "zil suspending";
3888 
3889 /*
3890  * Suspend an intent log.  While in suspended mode, we still honor
3891  * synchronous semantics, but we rely on txg_wait_synced() to do it.
3892  * On old version pools, we suspend the log briefly when taking a
3893  * snapshot so that it will have an empty intent log.
3894  *
3895  * Long holds are not really intended to be used the way we do here --
3896  * held for such a short time.  A concurrent caller of dsl_dataset_long_held()
3897  * could fail.  Therefore we take pains to only put a long hold if it is
3898  * actually necessary.  Fortunately, it will only be necessary if the
3899  * objset is currently mounted (or the ZVOL equivalent).  In that case it
3900  * will already have a long hold, so we are not really making things any worse.
3901  *
3902  * Ideally, we would locate the existing long-holder (i.e. the zfsvfs_t or
3903  * zvol_state_t), and use their mechanism to prevent their hold from being
3904  * dropped (e.g. VFS_HOLD()).  However, that would be even more pain for
3905  * very little gain.
3906  *
3907  * if cookiep == NULL, this does both the suspend & resume.
3908  * Otherwise, it returns with the dataset "long held", and the cookie
3909  * should be passed into zil_resume().
3910  */
3911 int
3912 zil_suspend(const char *osname, void **cookiep)
3913 {
3914 	objset_t *os;
3915 	zilog_t *zilog;
3916 	const zil_header_t *zh;
3917 	int error;
3918 
3919 	error = dmu_objset_hold(osname, suspend_tag, &os);
3920 	if (error != 0)
3921 		return (error);
3922 	zilog = dmu_objset_zil(os);
3923 
3924 	mutex_enter(&zilog->zl_issuer_lock);
3925 	mutex_enter(&zilog->zl_lock);
3926 	zh = zilog->zl_header;
3927 
3928 	if (zh->zh_flags & ZIL_REPLAY_NEEDED) {		/* unplayed log */
3929 		mutex_exit(&zilog->zl_lock);
3930 		mutex_exit(&zilog->zl_issuer_lock);
3931 		dmu_objset_rele(os, suspend_tag);
3932 		return (SET_ERROR(EBUSY));
3933 	}
3934 
3935 	/*
3936 	 * Don't put a long hold in the cases where we can avoid it.  This
3937 	 * is when there is no cookie so we are doing a suspend & resume
3938 	 * (i.e. called from zil_vdev_offline()), and there's nothing to do
3939 	 * for the suspend because it's already suspended, or there's no ZIL.
3940 	 */
3941 	if (cookiep == NULL && !zilog->zl_suspending &&
3942 	    (zilog->zl_suspend > 0 || BP_IS_HOLE(&zh->zh_log))) {
3943 		mutex_exit(&zilog->zl_lock);
3944 		mutex_exit(&zilog->zl_issuer_lock);
3945 		dmu_objset_rele(os, suspend_tag);
3946 		return (0);
3947 	}
3948 
3949 	dsl_dataset_long_hold(dmu_objset_ds(os), suspend_tag);
3950 	dsl_pool_rele(dmu_objset_pool(os), suspend_tag);
3951 
3952 	zilog->zl_suspend++;
3953 	mutex_exit(&zilog->zl_issuer_lock);
3954 
3955 	if (zilog->zl_suspend > 1) {
3956 		/*
3957 		 * Someone else is already suspending it.
3958 		 * Just wait for them to finish.
3959 		 */
3960 
3961 		while (zilog->zl_suspending)
3962 			cv_wait(&zilog->zl_cv_suspend, &zilog->zl_lock);
3963 		mutex_exit(&zilog->zl_lock);
3964 
3965 		if (cookiep == NULL)
3966 			zil_resume(os);
3967 		else
3968 			*cookiep = os;
3969 		return (0);
3970 	}
3971 
3972 	/*
3973 	 * If there is no pointer to an on-disk block, this ZIL must not
3974 	 * be active (e.g. filesystem not mounted), so there's nothing
3975 	 * to clean up.
3976 	 */
3977 	if (BP_IS_HOLE(&zh->zh_log)) {
3978 		ASSERT(cookiep != NULL); /* fast path already handled */
3979 
3980 		*cookiep = os;
3981 		mutex_exit(&zilog->zl_lock);
3982 		return (0);
3983 	}
3984 
3985 	/*
3986 	 * The ZIL has work to do. Ensure that the associated encryption
3987 	 * key will remain mapped while we are committing the log by
3988 	 * grabbing a reference to it. If the key isn't loaded we have no
3989 	 * choice but to return an error until the wrapping key is loaded.
3990 	 */
3991 	if (os->os_encrypted &&
3992 	    dsl_dataset_create_key_mapping(dmu_objset_ds(os)) != 0) {
3993 		zilog->zl_suspend--;
3994 		mutex_exit(&zilog->zl_lock);
3995 		dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
3996 		dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
3997 		return (SET_ERROR(EACCES));
3998 	}
3999 
4000 	zilog->zl_suspending = B_TRUE;
4001 	mutex_exit(&zilog->zl_lock);
4002 
4003 	/*
4004 	 * We need to use zil_commit_impl to ensure we wait for all
4005 	 * LWB_STATE_OPENED and LWB_STATE_ISSUED lwbs to be committed
4006 	 * to disk before proceeding. If we used zil_commit instead, it
4007 	 * would just call txg_wait_synced(), because zl_suspend is set.
4008 	 * txg_wait_synced() doesn't wait for these lwb's to be
4009 	 * LWB_STATE_FLUSH_DONE before returning.
4010 	 */
4011 	zil_commit_impl(zilog, 0);
4012 
4013 	/*
4014 	 * Now that we've ensured all lwb's are LWB_STATE_FLUSH_DONE, we
4015 	 * use txg_wait_synced() to ensure the data from the zilog has
4016 	 * migrated to the main pool before calling zil_destroy().
4017 	 */
4018 	txg_wait_synced(zilog->zl_dmu_pool, 0);
4019 
4020 	zil_destroy(zilog, B_FALSE);
4021 
4022 	mutex_enter(&zilog->zl_lock);
4023 	zilog->zl_suspending = B_FALSE;
4024 	cv_broadcast(&zilog->zl_cv_suspend);
4025 	mutex_exit(&zilog->zl_lock);
4026 
4027 	if (os->os_encrypted)
4028 		dsl_dataset_remove_key_mapping(dmu_objset_ds(os));
4029 
4030 	if (cookiep == NULL)
4031 		zil_resume(os);
4032 	else
4033 		*cookiep = os;
4034 	return (0);
4035 }
4036 
4037 void
4038 zil_resume(void *cookie)
4039 {
4040 	objset_t *os = cookie;
4041 	zilog_t *zilog = dmu_objset_zil(os);
4042 
4043 	mutex_enter(&zilog->zl_lock);
4044 	ASSERT(zilog->zl_suspend != 0);
4045 	zilog->zl_suspend--;
4046 	mutex_exit(&zilog->zl_lock);
4047 	dsl_dataset_long_rele(dmu_objset_ds(os), suspend_tag);
4048 	dsl_dataset_rele(dmu_objset_ds(os), suspend_tag);
4049 }
4050 
4051 typedef struct zil_replay_arg {
4052 	zil_replay_func_t *const *zr_replay;
4053 	void		*zr_arg;
4054 	boolean_t	zr_byteswap;
4055 	char		*zr_lr;
4056 } zil_replay_arg_t;
4057 
4058 static int
4059 zil_replay_error(zilog_t *zilog, const lr_t *lr, int error)
4060 {
4061 	char name[ZFS_MAX_DATASET_NAME_LEN];
4062 
4063 	zilog->zl_replaying_seq--;	/* didn't actually replay this one */
4064 
4065 	dmu_objset_name(zilog->zl_os, name);
4066 
4067 	cmn_err(CE_WARN, "ZFS replay transaction error %d, "
4068 	    "dataset %s, seq 0x%llx, txtype %llu %s\n", error, name,
4069 	    (u_longlong_t)lr->lrc_seq,
4070 	    (u_longlong_t)(lr->lrc_txtype & ~TX_CI),
4071 	    (lr->lrc_txtype & TX_CI) ? "CI" : "");
4072 
4073 	return (error);
4074 }
4075 
4076 static int
4077 zil_replay_log_record(zilog_t *zilog, const lr_t *lr, void *zra,
4078     uint64_t claim_txg)
4079 {
4080 	zil_replay_arg_t *zr = zra;
4081 	const zil_header_t *zh = zilog->zl_header;
4082 	uint64_t reclen = lr->lrc_reclen;
4083 	uint64_t txtype = lr->lrc_txtype;
4084 	int error = 0;
4085 
4086 	zilog->zl_replaying_seq = lr->lrc_seq;
4087 
4088 	if (lr->lrc_seq <= zh->zh_replay_seq)	/* already replayed */
4089 		return (0);
4090 
4091 	if (lr->lrc_txg < claim_txg)		/* already committed */
4092 		return (0);
4093 
4094 	/* Strip case-insensitive bit, still present in log record */
4095 	txtype &= ~TX_CI;
4096 
4097 	if (txtype == 0 || txtype >= TX_MAX_TYPE)
4098 		return (zil_replay_error(zilog, lr, EINVAL));
4099 
4100 	/*
4101 	 * If this record type can be logged out of order, the object
4102 	 * (lr_foid) may no longer exist.  That's legitimate, not an error.
4103 	 */
4104 	if (TX_OOO(txtype)) {
4105 		error = dmu_object_info(zilog->zl_os,
4106 		    LR_FOID_GET_OBJ(((lr_ooo_t *)lr)->lr_foid), NULL);
4107 		if (error == ENOENT || error == EEXIST)
4108 			return (0);
4109 	}
4110 
4111 	/*
4112 	 * Make a copy of the data so we can revise and extend it.
4113 	 */
4114 	memcpy(zr->zr_lr, lr, reclen);
4115 
4116 	/*
4117 	 * If this is a TX_WRITE with a blkptr, suck in the data.
4118 	 */
4119 	if (txtype == TX_WRITE && reclen == sizeof (lr_write_t)) {
4120 		error = zil_read_log_data(zilog, (lr_write_t *)lr,
4121 		    zr->zr_lr + reclen);
4122 		if (error != 0)
4123 			return (zil_replay_error(zilog, lr, error));
4124 	}
4125 
4126 	/*
4127 	 * The log block containing this lr may have been byteswapped
4128 	 * so that we can easily examine common fields like lrc_txtype.
4129 	 * However, the log is a mix of different record types, and only the
4130 	 * replay vectors know how to byteswap their records.  Therefore, if
4131 	 * the lr was byteswapped, undo it before invoking the replay vector.
4132 	 */
4133 	if (zr->zr_byteswap)
4134 		byteswap_uint64_array(zr->zr_lr, reclen);
4135 
4136 	/*
4137 	 * We must now do two things atomically: replay this log record,
4138 	 * and update the log header sequence number to reflect the fact that
4139 	 * we did so. At the end of each replay function the sequence number
4140 	 * is updated if we are in replay mode.
4141 	 */
4142 	error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, zr->zr_byteswap);
4143 	if (error != 0) {
4144 		/*
4145 		 * The DMU's dnode layer doesn't see removes until the txg
4146 		 * commits, so a subsequent claim can spuriously fail with
4147 		 * EEXIST. So if we receive any error we try syncing out
4148 		 * any removes then retry the transaction.  Note that we
4149 		 * specify B_FALSE for byteswap now, so we don't do it twice.
4150 		 */
4151 		txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
4152 		error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lr, B_FALSE);
4153 		if (error != 0)
4154 			return (zil_replay_error(zilog, lr, error));
4155 	}
4156 	return (0);
4157 }
4158 
4159 static int
4160 zil_incr_blks(zilog_t *zilog, const blkptr_t *bp, void *arg, uint64_t claim_txg)
4161 {
4162 	(void) bp, (void) arg, (void) claim_txg;
4163 
4164 	zilog->zl_replay_blks++;
4165 
4166 	return (0);
4167 }
4168 
4169 /*
4170  * If this dataset has a non-empty intent log, replay it and destroy it.
4171  * Return B_TRUE if there were any entries to replay.
4172  */
4173 boolean_t
4174 zil_replay(objset_t *os, void *arg,
4175     zil_replay_func_t *const replay_func[TX_MAX_TYPE])
4176 {
4177 	zilog_t *zilog = dmu_objset_zil(os);
4178 	const zil_header_t *zh = zilog->zl_header;
4179 	zil_replay_arg_t zr;
4180 
4181 	if ((zh->zh_flags & ZIL_REPLAY_NEEDED) == 0) {
4182 		return (zil_destroy(zilog, B_TRUE));
4183 	}
4184 
4185 	zr.zr_replay = replay_func;
4186 	zr.zr_arg = arg;
4187 	zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
4188 	zr.zr_lr = vmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
4189 
4190 	/*
4191 	 * Wait for in-progress removes to sync before starting replay.
4192 	 */
4193 	txg_wait_synced(zilog->zl_dmu_pool, 0);
4194 
4195 	zilog->zl_replay = B_TRUE;
4196 	zilog->zl_replay_time = ddi_get_lbolt();
4197 	ASSERT(zilog->zl_replay_blks == 0);
4198 	(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
4199 	    zh->zh_claim_txg, B_TRUE);
4200 	vmem_free(zr.zr_lr, 2 * SPA_MAXBLOCKSIZE);
4201 
4202 	zil_destroy(zilog, B_FALSE);
4203 	txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
4204 	zilog->zl_replay = B_FALSE;
4205 
4206 	return (B_TRUE);
4207 }
4208 
4209 boolean_t
4210 zil_replaying(zilog_t *zilog, dmu_tx_t *tx)
4211 {
4212 	if (zilog->zl_sync == ZFS_SYNC_DISABLED)
4213 		return (B_TRUE);
4214 
4215 	if (zilog->zl_replay) {
4216 		dsl_dataset_dirty(dmu_objset_ds(zilog->zl_os), tx);
4217 		zilog->zl_replayed_seq[dmu_tx_get_txg(tx) & TXG_MASK] =
4218 		    zilog->zl_replaying_seq;
4219 		return (B_TRUE);
4220 	}
4221 
4222 	return (B_FALSE);
4223 }
4224 
4225 int
4226 zil_reset(const char *osname, void *arg)
4227 {
4228 	(void) arg;
4229 
4230 	int error = zil_suspend(osname, NULL);
4231 	/* EACCES means crypto key not loaded */
4232 	if ((error == EACCES) || (error == EBUSY))
4233 		return (SET_ERROR(error));
4234 	if (error != 0)
4235 		return (SET_ERROR(EEXIST));
4236 	return (0);
4237 }
4238 
4239 EXPORT_SYMBOL(zil_alloc);
4240 EXPORT_SYMBOL(zil_free);
4241 EXPORT_SYMBOL(zil_open);
4242 EXPORT_SYMBOL(zil_close);
4243 EXPORT_SYMBOL(zil_replay);
4244 EXPORT_SYMBOL(zil_replaying);
4245 EXPORT_SYMBOL(zil_destroy);
4246 EXPORT_SYMBOL(zil_destroy_sync);
4247 EXPORT_SYMBOL(zil_itx_create);
4248 EXPORT_SYMBOL(zil_itx_destroy);
4249 EXPORT_SYMBOL(zil_itx_assign);
4250 EXPORT_SYMBOL(zil_commit);
4251 EXPORT_SYMBOL(zil_claim);
4252 EXPORT_SYMBOL(zil_check_log_chain);
4253 EXPORT_SYMBOL(zil_sync);
4254 EXPORT_SYMBOL(zil_clean);
4255 EXPORT_SYMBOL(zil_suspend);
4256 EXPORT_SYMBOL(zil_resume);
4257 EXPORT_SYMBOL(zil_lwb_add_block);
4258 EXPORT_SYMBOL(zil_bp_tree_add);
4259 EXPORT_SYMBOL(zil_set_sync);
4260 EXPORT_SYMBOL(zil_set_logbias);
4261 EXPORT_SYMBOL(zil_sums_init);
4262 EXPORT_SYMBOL(zil_sums_fini);
4263 EXPORT_SYMBOL(zil_kstat_values_update);
4264 
4265 ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, UINT, ZMOD_RW,
4266 	"ZIL block open timeout percentage");
4267 
4268 ZFS_MODULE_PARAM(zfs_zil, zil_, min_commit_timeout, U64, ZMOD_RW,
4269 	"Minimum delay we care for ZIL block commit");
4270 
4271 ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW,
4272 	"Disable intent logging replay");
4273 
4274 ZFS_MODULE_PARAM(zfs_zil, zil_, nocacheflush, INT, ZMOD_RW,
4275 	"Disable ZIL cache flushes");
4276 
4277 ZFS_MODULE_PARAM(zfs_zil, zil_, slog_bulk, U64, ZMOD_RW,
4278 	"Limit in bytes slog sync writes per commit");
4279 
4280 ZFS_MODULE_PARAM(zfs_zil, zil_, maxblocksize, UINT, ZMOD_RW,
4281 	"Limit in bytes of ZIL log block size");
4282